diff --git a/config_files/model_tests/base_210M.yaml b/config_files/model_tests/base_210M.yaml
new file mode 100644
index 000000000..384d5fe29
--- /dev/null
+++ b/config_files/model_tests/base_210M.yaml
@@ -0,0 +1,2732 @@
+# =============================================================================
+# ~210M NON-EMBEDDING baseline (adapted from your 124M-class config)
+# Architecture changes vs. original:  n_layer 12->16, n_embd 768->1024, n_head 12->16, ffn_hidden 3072->4096
+#   -> 205.6M non-embedding params, 308.6M total, head_dim=64, aspect ratio d/L=64 (canonical scaled-GPT-2)
+# All widths kept multiple of 128 and head_dim fixed at 64 (Flash fast path + qk_norm stays 64).
+# See accompanying notes for LR/batch/token-budget rationale and ablation knobs.
+# =============================================================================
+settings:
+  experiment_id: ${modalities_env:experiment_id}
+  config_file_path: ${modalities_env:config_file_path}
+  referencing_keys:
+    sample_key: input_ids
+    target_key: target_ids
+    prediction_key: logits
+  cuda_env:
+    local_rank: ${cuda_env:LOCAL_RANK}
+    global_rank: ${cuda_env:RANK}
+    world_size: ${cuda_env:WORLD_SIZE}
+  paths:
+    checkpoint_saving_path: /raid/s3/opengptx/mfrey/tests/checkpoints
+    experiments_root_path: /raid/s3/opengptx/mfrey/tests/checkpoints
+  intervals:
+    training_log_interval_in_steps: 1
+    checkpointing_interval_in_steps: 25
+    evaluation_interval_in_steps: 25
+  consistency_enforcement:
+    enforce_tokens_per_step_consistency: false
+    enforce_last_step_logged: false
+    enforce_last_step_evaluated: false
+    enforce_last_step_checkpointed: false
+  step_profile:
+    gradient_accumulation_steps: 16
+    local_train_micro_batch_size: 8
+    sequence_length: 4096
+    #     16 GPUs -> mbs=4, gas=2
+    #     32 GPUs -> mbs=4, gas=1
+    #     64 GPUs -> mbs=2, gas=1
+    #      1 GPU  -> mbs=2, gas=64
+    dp_degree:
+      instance_key: dp_degree
+      pass_type: BY_REFERENCE
+  training_target:
+    # 20B tokens (~65x total params / ~97x non-embedding).
+    num_target_tokens: 20971520000 # 40000 steps * 524288 tokens/step
+    num_target_steps:
+      component_key: number_conversion
+      variant_key: num_steps_from_num_tokens
+      config:
+        dp_degree:
+          instance_key: dp_degree
+          pass_type: BY_REFERENCE
+        local_micro_batch_size: ${settings.step_profile.local_train_micro_batch_size}
+        global_num_tokens: ${settings.training_target.num_target_tokens}
+        sequence_length: ${settings.step_profile.sequence_length}
+        gradient_accumulation_steps: ${settings.step_profile.gradient_accumulation_steps}
+  training_progress:
+    global_num_seen_tokens: 0
+    num_seen_steps: 0
+    num_seen_samples: 0
+    last_step: -1
+
+collate_fn:
+  component_key: collate_fn
+  variant_key: gpt_2_llm_collator
+  config:
+    sample_key: ${settings.referencing_keys.sample_key}
+    target_key: ${settings.referencing_keys.target_key}
+
+train_dataset:
+  component_key: dataset
+  variant_key: combined
+  config:
+    datasets:
+    # --- score4.5+: 128 shards ---
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00000.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00001.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00002.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00003.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00004.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00005.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00006.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00007.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00008.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00009.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00010.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00011.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00012.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00013.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00014.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00015.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00016.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00017.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00018.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00019.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00020.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00021.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00022.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00023.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00024.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00025.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00026.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00027.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00028.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00029.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00030.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00031.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00032.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00033.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00034.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00035.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00036.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00037.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00038.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00039.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00040.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00041.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00042.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00043.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00044.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00045.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00046.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00047.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00048.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00049.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00050.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00051.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00052.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00053.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00054.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00055.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00056.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00057.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00058.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00059.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00060.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00061.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00062.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00063.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00064.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00065.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00066.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00067.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00068.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00069.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00070.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00071.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00072.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00073.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00074.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00075.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00076.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00077.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00078.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00079.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00080.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00081.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00082.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00083.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00084.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00085.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00086.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00087.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00088.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00089.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00090.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00091.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00092.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00093.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00094.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00095.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00096.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00097.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00098.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00099.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00100.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00101.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00102.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00103.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00104.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00105.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00106.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00107.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00108.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00109.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00110.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00111.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00112.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00113.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00114.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00115.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00116.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00117.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00118.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00119.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00120.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00121.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00122.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00123.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00124.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00125.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00126.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.5+/part-00127.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    # --- score4.0-4.5: 128 shards ---
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00000.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00001.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00002.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00003.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00004.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00005.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00006.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00007.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00008.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00009.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00010.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00011.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00012.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00013.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00014.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00015.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00016.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00017.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00018.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00019.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00020.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00021.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00022.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00023.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00024.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00025.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00026.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00027.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00028.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00029.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00030.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00031.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00032.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00033.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00034.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00035.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00036.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00037.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00038.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00039.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00040.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00041.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00042.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00043.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00044.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00045.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00046.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00047.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00048.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00049.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00050.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00051.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00052.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00053.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00054.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00055.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00056.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00057.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00058.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00059.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00060.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00061.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00062.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00063.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00064.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00065.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00066.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00067.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00068.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00069.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00070.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00071.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00072.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00073.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00074.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00075.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00076.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00077.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00078.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00079.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00080.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00081.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00082.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00083.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00084.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00085.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00086.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00087.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00088.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00089.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00090.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00091.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00092.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00093.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00094.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00095.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00096.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00097.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00098.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00099.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00100.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00101.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00102.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00103.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00104.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00105.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00106.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00107.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00108.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00109.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00110.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00111.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00112.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00113.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00114.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00115.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00116.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00117.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00118.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00119.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00120.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00121.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00122.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00123.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00124.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00125.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00126.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00127.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    # --- score3.5-4.0: 128 shards ---
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00000.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00001.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00002.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00003.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00004.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00005.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00006.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00007.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00008.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00009.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00010.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00011.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00012.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00013.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00014.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00015.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00016.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00017.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00018.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00019.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00020.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00021.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00022.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00023.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00024.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00025.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00026.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00027.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00028.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00029.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00030.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00031.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00032.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00033.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00034.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00035.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00036.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00037.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00038.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00039.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00040.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00041.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00042.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00043.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00044.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00045.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00046.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00047.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00048.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00049.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00050.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00051.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00052.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00053.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00054.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00055.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00056.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00057.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00058.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00059.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00060.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00061.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00062.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00063.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00064.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00065.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00066.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00067.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00068.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00069.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00070.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00071.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00072.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00073.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00074.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00075.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00076.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00077.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00078.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00079.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00080.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00081.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00082.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00083.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00084.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00085.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00086.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00087.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00088.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00089.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00090.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00091.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00092.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00093.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00094.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00095.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00096.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00097.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00098.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00099.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00100.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00101.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00102.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00103.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00104.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00105.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00106.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00107.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00108.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00109.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00110.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00111.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00112.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00113.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00114.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00115.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00116.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00117.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00118.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00119.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00120.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00121.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00122.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00123.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00124.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00125.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00126.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score3.5-4.0/part-00127.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+train_dataloader:
+  component_key: data_loader
+  variant_key: default
+  config:
+    num_workers: 2
+    pin_memory: true
+    dataloader_tag: train
+    dataset:
+      instance_key: train_dataset
+      pass_type: BY_REFERENCE
+    batch_sampler:
+      component_key: batch_sampler
+      variant_key: default
+      config:
+        batch_size: ${settings.step_profile.local_train_micro_batch_size}
+        drop_last: true
+        sampler:
+          component_key: sampler
+          variant_key: resumable_distributed_sampler
+          config:
+            dataset:
+              instance_key: train_dataset
+              pass_type: BY_REFERENCE
+            rank: ${settings.cuda_env.global_rank}
+            num_replicas: ${settings.cuda_env.world_size}
+            shuffle: true
+            seed: 42
+            drop_last: true
+            skip_num_global_samples: ${settings.training_progress.num_seen_samples}
+    collate_fn:
+      instance_key: collate_fn
+      pass_type: BY_REFERENCE
+eval_dataset:
+  component_key: dataset
+  variant_key: packed_mem_map_dataset_continuous
+  config:
+    raw_data_path: /raid/s3/opengptx/mehdi-ali/git_repos/experimentation/data/eval_data/validation.pbin
+    sequence_length: ${settings.step_profile.sequence_length}
+    sample_key: ${settings.referencing_keys.sample_key}
+eval_dataloader:
+  component_key: data_loader
+  variant_key: default
+  config:
+    num_workers: 2
+    pin_memory: true
+    dataloader_tag: eval
+    dataset:
+      instance_key: eval_dataset
+      pass_type: BY_REFERENCE
+    batch_sampler:
+      component_key: batch_sampler
+      variant_key: default
+      config:
+        batch_size: 2
+        drop_last: true
+        sampler:
+          component_key: sampler
+          variant_key: distributed_sampler
+          config:
+            rank: ${settings.cuda_env.global_rank}
+            num_replicas: ${settings.cuda_env.world_size}
+            shuffle: false
+            drop_last: true
+            dataset:
+              instance_key: eval_dataset
+              pass_type: BY_REFERENCE
+    collate_fn:
+      instance_key: collate_fn
+      pass_type: BY_REFERENCE
+eval_dataloaders:
+- instance_key: eval_dataloader
+  pass_type: BY_REFERENCE
+
+checkpoint_saving:
+  component_key: checkpoint_saving
+  variant_key: default
+  config:
+    checkpoint_saving_strategy:
+      component_key: checkpoint_saving_strategy
+      variant_key: save_k_most_recent_checkpoints_strategy
+      config:
+        k: -1
+    checkpoint_saving_execution:
+      component_key: checkpoint_saving_execution
+      variant_key: dcp
+      config:
+        checkpoint_path: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+        global_rank: ${settings.cuda_env.global_rank}
+        experiment_id: ${settings.experiment_id}
+
+loss_fn:
+  component_key: loss
+  variant_key: clm_cross_entropy_loss
+  config:
+    target_key: ${settings.referencing_keys.target_key}
+    prediction_key: ${settings.referencing_keys.prediction_key}
+
+device_mesh:
+  component_key: device_mesh
+  variant_key: default
+  config:
+    device_type: cuda
+    data_parallel_replicate_degree: 1
+    data_parallel_shard_degree: -1
+    world_size: ${settings.cuda_env.world_size}
+
+dp_degree:
+  component_key: number_conversion
+  variant_key: parallel_degree
+  config:
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+    parallelism_methods: [dp_shard, dp_replicate]
+
+app_state:
+  component_key: app_state
+  variant_key: raw
+  config:
+    model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    optimizer:
+      instance_key: optimizer
+      pass_type: BY_REFERENCE
+    lr_scheduler:
+      instance_key: lr_scheduler
+      pass_type: BY_REFERENCE
+
+initialized_model:
+  component_key: model
+  variant_key: model_initialized
+  config:
+    model:
+      instance_key: fsdp_model
+      pass_type: BY_REFERENCE
+    model_initializer:
+      component_key: model_initialization
+      variant_key: composed
+      config:
+        model_type: gpt2
+        weight_init_type: scaled
+        mean: 0.0
+        std: 0.02
+        num_layers: ${model_raw.config.n_layer}
+        multi_device_generator_policy: error
+
+fsdp_model:
+  component_key: model
+  variant_key: fsdp2_wrapped
+  config:
+    model:
+      instance_key: compiled_model
+      pass_type: BY_REFERENCE
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+    mixed_precision_settings:
+      param_dtype: BF_16
+      reduce_dtype: BF_16
+    block_names: [GPT2Block]
+
+compiled_model:
+  component_key: model
+  variant_key: compiled
+  config:
+    model:
+      instance_key: model_raw
+      pass_type: BY_REFERENCE
+    block_names: [GPT2Block]
+
+model_raw:
+  component_key: model
+  variant_key: gpt2
+  config:
+    use_meta_device: false
+    use_weight_tying: false
+    sample_key: ${settings.referencing_keys.sample_key}
+    poe_type: NOPE
+    sequence_length: ${settings.step_profile.sequence_length}
+    prediction_key: ${loss_fn.config.prediction_key}
+    vocab_size: 50304
+    # ---- ~210M NON-EMBEDDING shape ----
+    n_layer: 16
+    n_head_q: 16
+    n_head_kv: 16
+    ffn_hidden: 4096
+    n_embd: 1024
+    # ----------------------------------------------
+    dropout: 0.0
+    bias: false
+    attention_config:
+      qkv_transforms:
+        - type_hint: RotaryTransform
+          config:
+            n_embd: ${model_raw.config.n_embd}
+            n_head: ${model_raw.config.n_head_q}
+            seq_length_dim: -2
+            base_freq: 10000
+      qk_norm_config:
+        norm_type: pytorch_rms_norm
+        config:
+          normalized_shape: 64
+          eps: 1.0e-05
+    attention_implementation: pytorch_flash
+    activation_type: swiglu
+    attention_norm_config:
+      norm_type: pytorch_rms_norm
+      config:
+        normalized_shape: 1024
+        eps: 1.0e-05
+    ffn_norm_config:
+      norm_type: pytorch_rms_norm
+      config:
+        normalized_shape: 1024
+        eps: 1.0e-05
+    lm_head_norm_config:
+      norm_type: pytorch_rms_norm
+      config:
+        normalized_shape: 1024
+        eps: 1.0e-05
+
+lr_scheduler:
+  component_key: scheduler
+  variant_key: onecycle_lr
+  config:
+    optimizer:
+      instance_key: optimizer
+      pass_type: BY_REFERENCE
+    max_lr: 0.003
+    div_factor: 10          # initial LR = max_lr/10 = 3.0e-4
+    final_div_factor: 1     # final LR = initial = max_lr/10
+    total_steps: ${settings.training_target.num_target_steps}
+    pct_start: 0.01
+    anneal_strategy: cos
+    last_epoch: ${settings.training_progress.last_step}
+
+optimizer:
+  component_key: optimizer
+  variant_key: adam_w
+  config:
+    lr: 0.0003             # overridden by OneCycle; set to max_lr/div_factor for consistency
+    betas: [0.9, 0.95]
+    eps: 1e-8
+    weight_decay: 1e-1     # decoupled WD=0.1, standard (GPT-3/Llama/OLMo); also aids LR transfer across width
+    weight_decay_groups_excluded: [embedding, pytorch_rms_norm]
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+
+gradient_clipper:
+  component_key: gradient_clipper
+  variant_key: fsdp2
+  config:
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    norm_type: P2_NORM
+    max_norm: 1.0
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+
+progress_subscriber:
+  component_key: progress_subscriber
+  variant_key: rich
+  config:
+    global_rank: ${settings.cuda_env.global_rank}
+    num_seen_steps: ${settings.training_progress.num_seen_steps}
+    num_target_steps: ${settings.training_target.num_target_steps}
+    train_dataloader_tag: ${train_dataloader.config.dataloader_tag}
+    eval_dataloaders:
+      instance_key: eval_dataloaders
+      pass_type: BY_REFERENCE
+
+evaluation_subscriber:
+  component_key: results_subscriber
+  variant_key: wandb
+  config:
+    global_rank: ${settings.cuda_env.global_rank}
+    project: tests
+    mode: ONLINE
+    experiment_id: base_toadapt
+    directory: wandb_storage
+    config_file_path: ${settings.config_file_path}
+
+mfu_calculator:
+  component_key: mfu_calculator
+  variant_key: gpt2
+  config:
+    n_layer: ${model_raw.config.n_layer}
+    sequence_length: ${settings.step_profile.sequence_length}
+    n_embd: ${model_raw.config.n_embd}
+    world_size: ${settings.cuda_env.world_size}
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+
+tokenizer:
+  component_key: tokenizer
+  variant_key: pretrained_hf_tokenizer
+  config:
+    pretrained_model_name_or_path: gpt2
+
+model_converter:
+  component_key: model_converter
+  variant_key: default
+  config:
+    command_template: "CUDA_VISIBLE_DEVICES=7 python src/modalities/conversion/gpt2/convert_gpt2.py {modalities_config} {output_dir} --checkpoint_path {checkpoint_path} > {checkpoint_path}/conversion.log 2>&1"
+    checkpoint_dir: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+    global_rank: ${settings.cuda_env.global_rank}
+    eval_interval: ${settings.intervals.checkpointing_interval_in_steps}
+
+downstream_evaluator:
+  component_key: downstream_evaluator
+  variant_key: default
+  config:
+    tokenizer:
+      instance_key: tokenizer
+      pass_type: BY_REFERENCE
+    tasks:
+      - "minerva_math_algebra:bpb::olmes"
+      - "minerva_math_counting_and_probability:bpb::olmes"
+      - "minerva_math_geometry:bpb::olmes"
+      - "minerva_math_intermediate_algebra:bpb::olmes"
+      - "minerva_math_number_theory:bpb::olmes"
+      - "minerva_math_prealgebra:bpb::olmes"
+      - "minerva_math_precalculus:bpb::olmes"
+      - "arc_challenge:rc::olmes:full"
+      - "arc_easy:rc::olmes:full"
+      - "hellaswag:rc::olmes:full"
+      - "winogrande:rc::olmes:full"
+      - "socialiqa:rc::olmes:full"
+      - "piqa:rc::olmes:full"
+      - "qasper_yesno:rc::olmes"
+      - "lambada"
+      - "arc_challenge:rc:bpb::olmes:full"
+      - "arc_easy:rc:bpb::olmes:full"
+      - "hellaswag:rc:bpb::olmes:full"
+      - "winogrande:rc:bpb::olmes:full"
+      - "socialiqa:rc:bpb::olmes:full"
+      - "piqa:rc:bpb::olmes:full"
+      - "qasper_yesno:rc:bpb::olmes"
+      - "lambada:bpb"
+      - "gsm8k::olmes"
+    eval_interval: ${settings.intervals.evaluation_interval_in_steps}
+    checkpoint_dir: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+    global_rank: ${settings.cuda_env.global_rank}
+    olmes_command_template: "CUDA_VISIBLE_DEVICES=7 . /home/markus_frey/Github/olmes/.venv/bin/activate && olmes --model {hf_model_dir} --model-args '{{\"trust_remote_code\": true}}' --task {tasks} --limit 8 --output-dir {hf_model_dir}/olmes_eval_{step} > {hf_model_dir}/olmes_eval_{step}.log 2>&1"
\ No newline at end of file
diff --git a/config_files/model_tests/base_toadapt.yaml b/config_files/model_tests/base_toadapt.yaml
new file mode 100644
index 000000000..a7d659f63
--- /dev/null
+++ b/config_files/model_tests/base_toadapt.yaml
@@ -0,0 +1,419 @@
+settings:
+  experiment_id: ${modalities_env:experiment_id}
+  config_file_path: ${modalities_env:config_file_path}
+  referencing_keys:
+    sample_key: input_ids
+    target_key: target_ids
+    prediction_key: logits
+  cuda_env:
+    local_rank: ${cuda_env:LOCAL_RANK}
+    global_rank: ${cuda_env:RANK}
+    world_size: ${cuda_env:WORLD_SIZE}
+  paths:
+    checkpoint_saving_path: /raid/s3/opengptx/mfrey/tests/checkpoints
+    experiments_root_path: /raid/s3/opengptx/mfrey/tests/checkpoints
+  intervals:
+    training_log_interval_in_steps: 1
+    checkpointing_interval_in_steps: 100
+    evaluation_interval_in_steps: 100
+  consistency_enforcement:
+    enforce_tokens_per_step_consistency: false
+    enforce_last_step_logged: false
+    enforce_last_step_evaluated: false
+    enforce_last_step_checkpointed: false
+  step_profile:
+    gradient_accumulation_steps: 2
+    local_train_micro_batch_size: 4
+    sequence_length: 4096
+    dp_degree:
+      instance_key: dp_degree
+      pass_type: BY_REFERENCE
+  training_target:
+    num_target_tokens: 10000000
+    num_target_steps:
+      component_key: number_conversion
+      variant_key: num_steps_from_num_tokens
+      config:
+        dp_degree:
+          instance_key: dp_degree
+          pass_type: BY_REFERENCE
+        local_micro_batch_size: ${settings.step_profile.local_train_micro_batch_size}
+        global_num_tokens: ${settings.training_target.num_target_tokens}
+        sequence_length: ${settings.step_profile.sequence_length}
+        gradient_accumulation_steps: ${settings.step_profile.gradient_accumulation_steps}
+  training_progress:
+    global_num_seen_tokens: 0
+    num_seen_steps: 0
+    num_seen_samples: 0
+    last_step: -1
+
+collate_fn:
+  component_key: collate_fn
+  variant_key: gpt_2_llm_collator
+  config:
+    sample_key: ${settings.referencing_keys.sample_key}
+    target_key: ${settings.referencing_keys.target_key}
+
+train_dataset:
+  component_key: dataset
+  variant_key: combined
+  config:
+    datasets:
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00000.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00001.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00002.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00003.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /raid/s3/opengptx/mehdi-ali/data/smollm-corpus-fw-edu-dedup-jsonl-buckets-raw/tokenized/score4.0-4.5/part-00004.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+train_dataloader:
+  component_key: data_loader
+  variant_key: default
+  config:
+    num_workers: 2
+    pin_memory: true
+    dataloader_tag: train
+    dataset:
+      instance_key: train_dataset
+      pass_type: BY_REFERENCE
+    batch_sampler:
+      component_key: batch_sampler
+      variant_key: default
+      config:
+        batch_size: ${settings.step_profile.local_train_micro_batch_size}
+        drop_last: true
+        sampler:
+          component_key: sampler
+          variant_key: resumable_distributed_sampler
+          config:
+            dataset:
+              instance_key: train_dataset
+              pass_type: BY_REFERENCE
+            rank: ${settings.cuda_env.global_rank}
+            num_replicas: ${settings.cuda_env.world_size}
+            shuffle: true
+            seed: 42
+            drop_last: true
+            skip_num_global_samples: ${settings.training_progress.num_seen_samples}
+    collate_fn:
+      instance_key: collate_fn
+      pass_type: BY_REFERENCE
+eval_dataset:
+  component_key: dataset
+  variant_key: packed_mem_map_dataset_continuous
+  config:
+    raw_data_path: /raid/s3/opengptx/mehdi-ali/git_repos/experimentation/data/eval_data/validation.pbin
+    sequence_length: ${settings.step_profile.sequence_length}
+    sample_key: ${settings.referencing_keys.sample_key}
+eval_dataloader:
+  component_key: data_loader
+  variant_key: default
+  config:
+    num_workers: 2
+    pin_memory: true
+    dataloader_tag: eval
+    dataset:
+      instance_key: eval_dataset
+      pass_type: BY_REFERENCE
+    batch_sampler:
+      component_key: batch_sampler
+      variant_key: default
+      config:
+        batch_size: 2
+        drop_last: true
+        sampler:
+          component_key: sampler
+          variant_key: distributed_sampler
+          config:
+            rank: ${settings.cuda_env.global_rank}
+            num_replicas: ${settings.cuda_env.world_size}
+            shuffle: false
+            drop_last: true
+            dataset:
+              instance_key: eval_dataset
+              pass_type: BY_REFERENCE
+    collate_fn:
+      instance_key: collate_fn
+      pass_type: BY_REFERENCE
+eval_dataloaders:
+- instance_key: eval_dataloader
+  pass_type: BY_REFERENCE
+
+checkpoint_saving:
+  component_key: checkpoint_saving
+  variant_key: default
+  config:
+    checkpoint_saving_strategy:
+      component_key: checkpoint_saving_strategy
+      variant_key: save_k_most_recent_checkpoints_strategy
+      config:
+        k: -1
+    checkpoint_saving_execution:
+      component_key: checkpoint_saving_execution
+      variant_key: dcp
+      config:
+        checkpoint_path: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+        global_rank: ${settings.cuda_env.global_rank}
+        experiment_id: ${settings.experiment_id}
+
+loss_fn:
+  component_key: loss
+  variant_key: clm_cross_entropy_loss
+  config:
+    target_key: ${settings.referencing_keys.target_key}
+    prediction_key: ${settings.referencing_keys.prediction_key}
+
+device_mesh:
+  component_key: device_mesh
+  variant_key: default
+  config:
+    device_type: cuda
+    data_parallel_replicate_degree: 1
+    data_parallel_shard_degree: -1
+    world_size: ${settings.cuda_env.world_size}
+
+dp_degree:
+  component_key: number_conversion
+  variant_key: parallel_degree
+  config:
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+    parallelism_methods: [dp_shard, dp_replicate]
+
+app_state:
+  component_key: app_state
+  variant_key: raw
+  config:
+    model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    optimizer:
+      instance_key: optimizer
+      pass_type: BY_REFERENCE
+    lr_scheduler:
+      instance_key: lr_scheduler
+      pass_type: BY_REFERENCE
+
+initialized_model:
+  component_key: model
+  variant_key: model_initialized
+  config:
+    model:
+      instance_key: fsdp_model
+      pass_type: BY_REFERENCE
+    model_initializer:
+      component_key: model_initialization
+      variant_key: composed
+      config:
+        model_type: gpt2
+        weight_init_type: scaled
+        mean: 0.0
+        std: 0.02
+        num_layers: ${model_raw.config.n_layer}
+        multi_device_generator_policy: error
+
+fsdp_model:
+  component_key: model
+  variant_key: fsdp2_wrapped
+  config:
+    model:
+      instance_key: compiled_model
+      pass_type: BY_REFERENCE
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+    mixed_precision_settings:
+      param_dtype: BF_16
+      reduce_dtype: BF_16
+    block_names: [GPT2Block]
+
+compiled_model:
+  component_key: model
+  variant_key: compiled
+  config:
+    model:
+      instance_key: model_raw
+      pass_type: BY_REFERENCE
+    block_names: [GPT2Block]
+
+model_raw:
+  component_key: model
+  variant_key: gpt2
+  config:
+    use_meta_device: false
+    use_weight_tying: false
+    sample_key: ${settings.referencing_keys.sample_key}
+    poe_type: NOPE
+    sequence_length: ${settings.step_profile.sequence_length}
+    prediction_key: ${loss_fn.config.prediction_key}
+    vocab_size: 50304
+    n_layer: 12
+    n_head_q: 12
+    n_head_kv: 12
+    ffn_hidden: 3072
+    n_embd: 768
+    dropout: 0.0
+    bias: false
+    attention_config:
+      qkv_transforms:
+        - type_hint: RotaryTransform
+          config:
+            n_embd: ${model_raw.config.n_embd}
+            n_head: ${model_raw.config.n_head_q}
+            seq_length_dim: -2
+            base_freq: 500000
+      qk_norm_config:
+        norm_type: pytorch_rms_norm
+        config:
+          normalized_shape: 64 # n_embd / n_head_q
+          eps: 1.0e-05
+    attention_implementation: pytorch_flash
+    activation_type: swiglu
+    attention_norm_config:
+      norm_type: pytorch_rms_norm
+      config:
+        normalized_shape: 768
+        eps: 1.0e-05
+    ffn_norm_config:
+      norm_type: pytorch_rms_norm
+      config:
+        normalized_shape: 768
+        eps: 1.0e-05
+    lm_head_norm_config:
+      norm_type: pytorch_rms_norm
+      config:
+        normalized_shape: 768
+        eps: 1.0e-05
+
+lr_scheduler:
+  component_key: scheduler
+  variant_key: onecycle_lr
+  config:
+    optimizer:
+      instance_key: optimizer
+      pass_type: BY_REFERENCE
+    max_lr: 0.003
+    div_factor: 10
+    final_div_factor: 1
+    total_steps: ${settings.training_target.num_target_steps}
+    pct_start: 0.01
+    anneal_strategy: cos
+    last_epoch: ${settings.training_progress.last_step}
+
+optimizer:
+  component_key: optimizer
+  variant_key: adam_w
+  config:
+    lr: 0.0001
+    betas: [0.9, 0.95]
+    eps: 1e-8
+    weight_decay: 1e-1
+    weight_decay_groups_excluded: [embedding, pytorch_rms_norm]
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+
+gradient_clipper:
+  component_key: gradient_clipper
+  variant_key: fsdp2
+  config:
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    norm_type: P2_NORM
+    max_norm: 1.0
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+
+progress_subscriber:
+  component_key: progress_subscriber
+  variant_key: rich
+  config:
+    global_rank: ${settings.cuda_env.global_rank}
+    num_seen_steps: ${settings.training_progress.num_seen_steps}
+    num_target_steps: ${settings.training_target.num_target_steps}
+    train_dataloader_tag: ${train_dataloader.config.dataloader_tag}
+    eval_dataloaders:
+      instance_key: eval_dataloaders
+      pass_type: BY_REFERENCE
+
+evaluation_subscriber:
+  component_key: results_subscriber
+  variant_key: wandb
+  config:
+    global_rank: ${settings.cuda_env.global_rank}
+    project: tests
+    mode: ONLINE
+    experiment_id: base_toadapt
+    directory: wandb_storage
+    config_file_path: ${settings.config_file_path}
+
+mfu_calculator:
+  component_key: mfu_calculator
+  variant_key: gpt2
+  config:
+    n_layer: ${model_raw.config.n_layer}
+    sequence_length: ${settings.step_profile.sequence_length}
+    n_embd: ${model_raw.config.n_embd}
+    world_size: ${settings.cuda_env.world_size}
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+
+tokenizer:
+  component_key: tokenizer
+  variant_key: pretrained_hf_tokenizer
+  config:
+    pretrained_model_name_or_path: gpt2
+
+model_converter:
+  component_key: model_converter
+  variant_key: default
+  config:
+    command_template: "CUDA_VISIBLE_DEVICES=7 python src/modalities/conversion/gpt2/convert_gpt2.py {modalities_config} {output_dir} --checkpoint_path {checkpoint_path} > {checkpoint_path}/conversion.log 2>&1"
+    checkpoint_dir: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+    global_rank: ${settings.cuda_env.global_rank}
+    eval_interval: 100
+
+downstream_evaluator:
+  component_key: downstream_evaluator
+  variant_key: default
+  config:
+    tokenizer:
+      instance_key: tokenizer
+      pass_type: BY_REFERENCE
+    tasks:
+      - "arc_challenge::olmes"
+      - "hellaswag::olmes"
+    eval_interval: 100
+    checkpoint_dir: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+    global_rank: ${settings.cuda_env.global_rank}
+    olmes_command_template: "CUDA_VISIBLE_DEVICES=7 . /home/markus_frey/Github/olmes/.venv/bin/activate && olmes --model {hf_model_dir} --model-args '{{\"trust_remote_code\": true}}' --task {tasks} --limit 128 --output-dir {hf_model_dir}/olmes_eval_{step} > {hf_model_dir}/olmes_eval_{step}.log 2>&1"
\ No newline at end of file
diff --git a/config_files/model_tests/leonardo/base_210M.yaml b/config_files/model_tests/leonardo/base_210M.yaml
new file mode 100644
index 000000000..a846e4255
--- /dev/null
+++ b/config_files/model_tests/leonardo/base_210M.yaml
@@ -0,0 +1,2728 @@
+# =============================================================================
+# ~210M NON-EMBEDDING baseline (adapted from your 124M-class config)
+# Architecture changes vs. original:  n_layer 12->16, n_embd 768->1024, n_head 12->16, ffn_hidden 3072->4096
+#   -> 205.6M non-embedding params, 308.6M total, head_dim=64, aspect ratio d/L=64 (canonical scaled-GPT-2)
+# All widths kept multiple of 128 and head_dim fixed at 64 (Flash fast path + qk_norm stays 64).
+# See accompanying notes for LR/batch/token-budget rationale and ablation knobs.
+# =============================================================================
+settings:
+  experiment_id: ${modalities_env:experiment_id}
+  config_file_path: ${modalities_env:config_file_path}
+  referencing_keys:
+    sample_key: input_ids
+    target_key: target_ids
+    prediction_key: logits
+  cuda_env:
+    local_rank: ${cuda_env:LOCAL_RANK}
+    global_rank: ${cuda_env:RANK}
+    world_size: ${cuda_env:WORLD_SIZE}
+  paths:
+    checkpoint_saving_path: /leonardo_scratch/large/userexternal/mfrey000/experiments_tests
+    experiments_root_path: /leonardo_scratch/large/userexternal/mfrey000/experiments_tests
+  intervals:
+    training_log_interval_in_steps: 1
+    checkpointing_interval_in_steps: 10000
+    evaluation_interval_in_steps: 10000
+  consistency_enforcement:
+    enforce_tokens_per_step_consistency: true
+    enforce_last_step_logged: false
+    enforce_last_step_evaluated: false
+    enforce_last_step_checkpointed: false
+  step_profile:
+    gradient_accumulation_steps: 1
+    local_train_micro_batch_size: 8
+    sequence_length: 4096
+    #     16 GPUs -> mbs=4, gas=2
+    #     32 GPUs -> mbs=4, gas=1
+    #     64 GPUs -> mbs=2, gas=1
+    #      1 GPU  -> mbs=2, gas=64
+    dp_degree:
+      instance_key: dp_degree
+      pass_type: BY_REFERENCE
+  training_target:
+    # 20B tokens (~65x total params / ~97x non-embedding).
+    num_target_tokens: 20971520000 # 40000 steps * 524288 tokens/step
+    num_target_steps:
+      component_key: number_conversion
+      variant_key: num_steps_from_num_tokens
+      config:
+        dp_degree:
+          instance_key: dp_degree
+          pass_type: BY_REFERENCE
+        local_micro_batch_size: ${settings.step_profile.local_train_micro_batch_size}
+        global_num_tokens: ${settings.training_target.num_target_tokens}
+        sequence_length: ${settings.step_profile.sequence_length}
+        gradient_accumulation_steps: ${settings.step_profile.gradient_accumulation_steps}
+  training_progress:
+    global_num_seen_tokens: 0
+    num_seen_steps: 0
+    num_seen_samples: 0
+    last_step: -1
+
+collate_fn:
+  component_key: collate_fn
+  variant_key: gpt_2_llm_collator
+  config:
+    sample_key: ${settings.referencing_keys.sample_key}
+    target_key: ${settings.referencing_keys.target_key}
+
+train_dataset:
+  component_key: dataset
+  variant_key: combined
+  config:
+    datasets:
+    # --- score4.5+: 128 shards ---
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00001.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00002.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00003.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00004.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00005.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00006.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00007.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00008.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00009.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00010.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00011.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00012.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00013.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00014.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00015.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00016.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00017.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00018.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00019.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00020.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00021.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00022.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00023.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00024.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00025.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00026.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00027.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00028.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00029.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00030.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00031.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00032.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00033.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00034.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00035.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00036.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00037.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00038.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00039.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00040.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00041.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00042.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00043.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00044.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00045.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00046.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00047.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00048.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00049.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00050.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00051.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00052.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00053.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00054.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00055.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00056.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00057.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00058.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00059.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00060.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00061.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00062.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00063.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00064.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00065.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00066.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00067.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00068.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00069.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00070.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00071.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00072.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00073.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00074.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00075.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00076.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00077.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00078.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00079.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00080.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00081.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00082.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00083.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00084.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00085.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00086.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00087.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00088.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00089.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00090.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00091.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00092.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00093.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00094.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00095.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00096.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00097.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00098.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00099.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00100.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00101.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00102.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00103.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00104.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00105.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00106.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00107.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00108.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00109.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00110.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00111.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00112.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00113.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00114.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00115.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00116.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00117.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00118.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00119.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00120.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00121.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00122.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00123.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00124.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00125.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00126.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00127.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    # --- score4.0-4.5: 128 shards ---
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00000.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00001.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00002.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00003.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00004.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00005.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00006.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00007.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00008.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00009.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00010.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00011.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00012.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00013.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00014.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00015.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00016.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00017.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00018.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00019.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00020.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00021.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00022.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00023.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00024.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00025.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00026.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00027.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00028.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00029.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00030.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00031.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00032.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00033.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00034.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00035.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00036.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00037.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00038.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00039.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00040.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00041.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00042.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00043.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00044.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00045.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00046.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00047.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00048.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00049.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00050.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00051.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00052.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00053.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00054.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00055.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00056.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00057.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00058.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00059.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00060.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00061.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00062.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00063.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00064.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00065.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00066.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00067.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00068.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00069.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00070.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00071.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00072.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00073.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00074.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00075.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00076.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00077.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00078.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00079.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00080.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00081.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00082.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00083.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00084.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00085.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00086.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00087.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00088.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00089.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00090.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00091.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00092.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00093.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00094.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00095.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00096.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00097.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00098.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00099.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00100.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00101.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00102.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00103.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00104.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00105.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00106.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00107.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00108.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00109.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00110.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00111.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00112.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00113.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00114.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00115.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00116.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00117.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00118.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00119.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00120.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00121.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00122.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00123.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00124.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00125.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00126.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.0-4.5/part-00127.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    # --- score3.5-4.0: 128 shards ---
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00000.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00001.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00002.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00003.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00004.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00005.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00006.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00007.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00008.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00009.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00010.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00011.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00012.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00013.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00014.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00015.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00016.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00017.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00018.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00019.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00020.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00021.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00022.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00023.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00024.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00025.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00026.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00027.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00028.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00029.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00030.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00031.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00032.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00033.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00034.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00035.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00036.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00037.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00038.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00039.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00040.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00041.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00042.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00043.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00044.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00045.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00046.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00047.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00048.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00049.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00050.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00051.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00052.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00053.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00054.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00055.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00056.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00057.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00058.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00059.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00060.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00061.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00062.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00063.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00064.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00065.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00066.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00067.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00068.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00069.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00070.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00071.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00072.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00073.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00074.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00075.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00076.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00077.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00078.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00079.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00080.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00081.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00082.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00083.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00084.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00085.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00086.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00087.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00088.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00089.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00090.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00091.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00092.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00093.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00094.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00095.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00096.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00097.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00098.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00099.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00100.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00101.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00102.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00103.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00104.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00105.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00106.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00107.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00108.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00109.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00110.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00111.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00112.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00113.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00114.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00115.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00116.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00117.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00118.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00119.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00120.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00121.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00122.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00123.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00124.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00125.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00126.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+    - component_key: dataset
+      variant_key: packed_mem_map_dataset_continuous
+      config:
+        raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score3.5-4.0/part-00127.pbin
+        sequence_length: ${settings.step_profile.sequence_length}
+        sample_key: ${settings.referencing_keys.sample_key}
+train_dataloader:
+  component_key: data_loader
+  variant_key: default
+  config:
+    num_workers: 2
+    pin_memory: true
+    dataloader_tag: train
+    dataset:
+      instance_key: train_dataset
+      pass_type: BY_REFERENCE
+    batch_sampler:
+      component_key: batch_sampler
+      variant_key: default
+      config:
+        batch_size: ${settings.step_profile.local_train_micro_batch_size}
+        drop_last: true
+        sampler:
+          component_key: sampler
+          variant_key: resumable_distributed_sampler
+          config:
+            dataset:
+              instance_key: train_dataset
+              pass_type: BY_REFERENCE
+            rank: ${settings.cuda_env.global_rank}
+            num_replicas: ${settings.cuda_env.world_size}
+            shuffle: true
+            seed: 42
+            drop_last: true
+            skip_num_global_samples: ${settings.training_progress.num_seen_samples}
+    collate_fn:
+      instance_key: collate_fn
+      pass_type: BY_REFERENCE
+eval_dataset:
+  component_key: dataset
+  variant_key: packed_mem_map_dataset_continuous
+  config:
+    raw_data_path: /leonardo_work/EUHPC_E05_119/mfrey/tokenized/score4.5+/part-00000.pbin
+    sequence_length: ${settings.step_profile.sequence_length}
+    sample_key: ${settings.referencing_keys.sample_key}
+eval_dataloader:
+  component_key: data_loader
+  variant_key: default
+  config:
+    num_workers: 2
+    pin_memory: true
+    dataloader_tag: eval
+    dataset:
+      instance_key: eval_dataset
+      pass_type: BY_REFERENCE
+    batch_sampler:
+      component_key: batch_sampler
+      variant_key: default
+      config:
+        batch_size: 2
+        drop_last: true
+        sampler:
+          component_key: sampler
+          variant_key: distributed_sampler
+          config:
+            rank: ${settings.cuda_env.global_rank}
+            num_replicas: ${settings.cuda_env.world_size}
+            shuffle: false
+            drop_last: true
+            dataset:
+              instance_key: eval_dataset
+              pass_type: BY_REFERENCE
+    collate_fn:
+      instance_key: collate_fn
+      pass_type: BY_REFERENCE
+eval_dataloaders:
+- instance_key: eval_dataloader
+  pass_type: BY_REFERENCE
+
+checkpoint_saving:
+  component_key: checkpoint_saving
+  variant_key: default
+  config:
+    checkpoint_saving_strategy:
+      component_key: checkpoint_saving_strategy
+      variant_key: save_k_most_recent_checkpoints_strategy
+      config:
+        k: -1
+    checkpoint_saving_execution:
+      component_key: checkpoint_saving_execution
+      variant_key: dcp
+      config:
+        checkpoint_path: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+        global_rank: ${settings.cuda_env.global_rank}
+        experiment_id: ${settings.experiment_id}
+
+loss_fn:
+  component_key: loss
+  variant_key: clm_cross_entropy_loss
+  config:
+    target_key: ${settings.referencing_keys.target_key}
+    prediction_key: ${settings.referencing_keys.prediction_key}
+
+device_mesh:
+  component_key: device_mesh
+  variant_key: default
+  config:
+    device_type: cuda
+    data_parallel_replicate_degree: 1
+    data_parallel_shard_degree: -1
+    world_size: ${settings.cuda_env.world_size}
+
+dp_degree:
+  component_key: number_conversion
+  variant_key: parallel_degree
+  config:
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+    parallelism_methods: [dp_shard, dp_replicate]
+
+app_state:
+  component_key: app_state
+  variant_key: raw
+  config:
+    model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    optimizer:
+      instance_key: optimizer
+      pass_type: BY_REFERENCE
+    lr_scheduler:
+      instance_key: lr_scheduler
+      pass_type: BY_REFERENCE
+
+initialized_model:
+  component_key: model
+  variant_key: model_initialized
+  config:
+    model:
+      instance_key: fsdp_model
+      pass_type: BY_REFERENCE
+    model_initializer:
+      component_key: model_initialization
+      variant_key: composed
+      config:
+        model_type: gpt2
+        weight_init_type: scaled
+        mean: 0.0
+        std: 0.02
+        num_layers: ${model_raw.config.n_layer}
+        multi_device_generator_policy: error
+
+fsdp_model:
+  component_key: model
+  variant_key: fsdp2_wrapped
+  config:
+    model:
+      instance_key: compiled_model
+      pass_type: BY_REFERENCE
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+    mixed_precision_settings:
+      param_dtype: BF_16
+      reduce_dtype: BF_16
+    block_names: [GPT2Block]
+
+compiled_model:
+  component_key: model
+  variant_key: compiled
+  config:
+    model:
+      instance_key: model_raw
+      pass_type: BY_REFERENCE
+    block_names: [GPT2Block]
+
+model_raw:
+  component_key: model
+  variant_key: gpt2
+  config:
+    use_meta_device: false
+    use_weight_tying: false
+    sample_key: ${settings.referencing_keys.sample_key}
+    poe_type: NOPE
+    sequence_length: ${settings.step_profile.sequence_length}
+    prediction_key: ${loss_fn.config.prediction_key}
+    vocab_size: 50304
+    # ---- ~210M NON-EMBEDDING shape ----
+    n_layer: 16
+    n_head_q: 16
+    n_head_kv: 16
+    ffn_hidden: 4096
+    n_embd: 1024
+    # ----------------------------------------------
+    dropout: 0.0
+    bias: false
+    attention_config:
+      qkv_transforms:
+        - type_hint: RotaryTransform
+          config:
+            n_embd: ${model_raw.config.n_embd}
+            n_head: ${model_raw.config.n_head_q}
+            seq_length_dim: -2
+            base_freq: 10000
+      qk_norm_config:
+        norm_type: pytorch_rms_norm
+        config:
+          normalized_shape: 64
+          eps: 1.0e-05
+    attention_implementation: pytorch_flash
+    activation_type: swiglu
+    attention_norm_config:
+      norm_type: pytorch_rms_norm
+      config:
+        normalized_shape: 1024
+        eps: 1.0e-05
+    ffn_norm_config:
+      norm_type: pytorch_rms_norm
+      config:
+        normalized_shape: 1024
+        eps: 1.0e-05
+    lm_head_norm_config:
+      norm_type: pytorch_rms_norm
+      config:
+        normalized_shape: 1024
+        eps: 1.0e-05
+
+lr_scheduler:
+  component_key: scheduler
+  variant_key: onecycle_lr
+  config:
+    optimizer:
+      instance_key: optimizer
+      pass_type: BY_REFERENCE
+    max_lr: 0.003
+    div_factor: 10          # initial LR = max_lr/10 = 3.0e-4
+    final_div_factor: 1     # final LR = initial = max_lr/10
+    total_steps: ${settings.training_target.num_target_steps}
+    pct_start: 0.01
+    anneal_strategy: cos
+    last_epoch: ${settings.training_progress.last_step}
+
+optimizer:
+  component_key: optimizer
+  variant_key: adam_w
+  config:
+    lr: 0.0003             # overridden by OneCycle; set to max_lr/div_factor for consistency
+    betas: [0.9, 0.95]
+    eps: 1e-8
+    weight_decay: 1e-1     # decoupled WD=0.1, standard (GPT-3/Llama/OLMo); also aids LR transfer across width
+    weight_decay_groups_excluded: [embedding, pytorch_rms_norm]
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+
+gradient_clipper:
+  component_key: gradient_clipper
+  variant_key: fsdp2
+  config:
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    norm_type: P2_NORM
+    max_norm: 1.0
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+
+progress_subscriber:
+  component_key: progress_subscriber
+  variant_key: rich
+  config:
+    global_rank: ${settings.cuda_env.global_rank}
+    num_seen_steps: ${settings.training_progress.num_seen_steps}
+    num_target_steps: ${settings.training_target.num_target_steps}
+    train_dataloader_tag: ${train_dataloader.config.dataloader_tag}
+    eval_dataloaders:
+      instance_key: eval_dataloaders
+      pass_type: BY_REFERENCE
+
+evaluation_subscriber:
+  component_key: results_subscriber
+  variant_key: wandb
+  config:
+    global_rank: ${settings.cuda_env.global_rank}
+    project: tests
+    mode: OFFLINE
+    experiment_id: base_toadapt
+    directory: ${settings.paths.experiments_root_path}/wandb
+    config_file_path: ${settings.config_file_path}
+
+mfu_calculator:
+  component_key: mfu_calculator
+  variant_key: gpt2
+  config:
+    n_layer: ${model_raw.config.n_layer}
+    sequence_length: ${settings.step_profile.sequence_length}
+    n_embd: ${model_raw.config.n_embd}
+    world_size: ${settings.cuda_env.world_size}
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+
+tokenizer:
+  component_key: tokenizer
+  variant_key: pretrained_hf_tokenizer
+  config:
+    pretrained_model_name_or_path: /leonardo_work/EUHPC_D21_101/mfrey/tokenizer/local_gpt2_tokenizer/
+    padding: false
+    truncation: false
+
+model_converter:
+  component_key: model_converter
+  variant_key: default
+  config:
+    command_template: "python src/modalities/conversion/gpt2/convert_gpt2.py {modalities_config} {output_dir} --checkpoint_path {checkpoint_path} > {checkpoint_path}/conversion.log 2>&1"
+    checkpoint_dir: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+    global_rank: ${settings.cuda_env.global_rank}
+    eval_interval: ${settings.intervals.checkpointing_interval_in_steps}
+
+downstream_evaluator:
+  component_key: downstream_evaluator
+  variant_key: default
+  config:
+    tokenizer:
+      instance_key: tokenizer
+      pass_type: BY_REFERENCE
+    tasks:
+      - "minerva_math_algebra:bpb::olmes"
+      - "minerva_math_counting_and_probability:bpb::olmes"
+      - "minerva_math_geometry:bpb::olmes"
+      - "minerva_math_intermediate_algebra:bpb::olmes"
+      - "minerva_math_number_theory:bpb::olmes"
+      - "minerva_math_prealgebra:bpb::olmes"
+      - "minerva_math_precalculus:bpb::olmes"
+      - "arc_challenge:rc::olmes:full"
+      - "arc_easy:rc::olmes:full"
+      - "hellaswag:rc::olmes:full"
+      - "winogrande:rc::olmes:full"
+      - "socialiqa:rc::olmes:full"
+      - "piqa:rc::olmes:full"
+      - "qasper_yesno:rc::olmes"
+      - "lambada"
+      - "arc_challenge:rc:bpb::olmes:full"
+      - "arc_easy:rc:bpb::olmes:full"
+      - "hellaswag:rc:bpb::olmes:full"
+      - "winogrande:rc:bpb::olmes:full"
+      - "socialiqa:rc:bpb::olmes:full"
+      - "piqa:rc:bpb::olmes:full"
+      - "qasper_yesno:rc:bpb::olmes"
+      - "lambada:bpb"
+      - "gsm8k::olmes"
+    eval_interval: ${settings.intervals.evaluation_interval_in_steps}
+    checkpoint_dir: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+    global_rank: ${settings.cuda_env.global_rank}
+    olmes_command_template: "bash scripts/evaluation/run_olmes_sbatch.sh {hf_model_dir} '{tasks}' {step} 128 1"
\ No newline at end of file
diff --git a/docs/components/downstream_evaluation.md b/docs/components/downstream_evaluation.md
new file mode 100644
index 000000000..8927ca16f
--- /dev/null
+++ b/docs/components/downstream_evaluation.md
@@ -0,0 +1,158 @@
+# Downstream Evaluation Pipeline
+
+## Overview
+
+The downstream evaluation pipeline in Modalities is a decoupled, three-stage callback system that executes at configurable step intervals during the training loop.
+
+The order of execution inside `Trainer.train` is:
+1. `checkpointing_callback`: Saves the PyTorch/FSDP checkpoint to disk.
+2. `conversion_callback`: (Optional) Converts the PyTorch checkpoint to a Hugging Face (HF) checkpoint.
+3. `downstream_evaluation_callback`: (Optional) Runs external evaluation tools (like OLMES) on the newly created HF checkpoint.
+
+By keeping conversion and evaluation decoupled, you can configure just the converter, just the evaluator (if HF checkpoints are generated elsewhere), or both.
+
+---
+
+## 1. Conversion Callback (`ModelConverter`)
+
+**Location:** `src/modalities/conversion/model_converter.py` (Lines 10-67)
+
+The `ModelConverter` is a thin wrapper that executes a shell command template via a subprocess.
+
+### Behavior
+- Triggered if `num_train_steps_done % eval_interval == 0`.
+- Only executes on `global_rank == 0`. You can prefix the command with `CUDA_VISIBLE_DEVICES=X` to manually specify which GPU the evaluation script should run on.
+- Reads `last_checkpoint_info.json` from the checkpoint directory to determine the latest checkpoint path.
+- Checks if the `{checkpoint_path}/hf_checkpoint` directory already exists. If it does, conversion is skipped.
+- If it does not exist, it formats the `command_template` and runs it using `subprocess.run(cmd, shell=True, check=True)`.
+
+### Placeholders
+The `command_template` string can use the following placeholders:
+- `{checkpoint_path}`: The path to the latest checkpoint directory (resolved at runtime).
+- `{output_dir}`: Evaluates to `{checkpoint_path}/hf_checkpoint`.
+- `{modalities_config}`: Path to the YAML config file found inside or next to the checkpoint directory.
+
+### YAML Configuration
+```yaml
+model_converter:
+  component_key: model_converter
+  variant_key: default
+  config:
+    command_template: "python src/modalities/conversion/gpt2/convert_gpt2.py {modalities_config} {output_dir} --checkpoint_path {checkpoint_path}"
+    checkpoint_dir: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+    global_rank: ${settings.cuda_env.global_rank}
+    eval_interval: 1000
+```
+
+---
+
+## 2. Downstream Evaluation Callback (`DownstreamEvaluator`)
+
+**Location:** `src/modalities/evaluator.py` (Lines 210-335)
+
+The `DownstreamEvaluator` checks for the existence of an HF checkpoint, launches an evaluation script via a subprocess, tracks active processes, and syncs OLMES metrics to the active W&B run.
+
+### Behavior
+- Triggered if `num_train_steps_done % eval_interval == 0`.
+- Only executes on `global_rank == 0`.
+- Reads `last_checkpoint_info.json` to find the latest checkpoint.
+- Checks if `{checkpoint_path}/hf_checkpoint` exists. If it does NOT exist, evaluation is skipped with a warning (assuming conversion failed or was disabled).
+- If the HF checkpoint exists, it formats the `olmes_command_template` and launches it asynchronously using `subprocess.Popen(cmd, shell=True)`.
+- **Process Tracking**: Stores `(Popen, step, hf_model_dir)` tuples in `self.active_processes` (Lines 233, 258).
+- **Graceful Exit**: `wait_for_evaluations()` (Lines 264-275) iterates over `active_processes`, calls `.wait()`, and syncs metrics after each evaluation completes.
+- **W&B Metric Sync**: `_sync_metrics_to_wandb()` (Lines 277-315) parses `metrics-all.jsonl` from the OLMES output directory, extracts `primary_score` for each task alias, and logs them to the active `wandb.run` as `eval/{alias}` at the correct training step. Gracefully skips if W&B is disabled or not installed.
+
+### Placeholders
+The `olmes_command_template` string can use the following placeholders:
+- `{hf_model_dir}`: The path to the `{checkpoint_path}/hf_checkpoint` directory.
+- `{tasks}`: A space-separated string of the tasks provided in the config (Line 248).
+- `{step}`: The current `num_train_steps_done`.
+
+### HPC / SLURM Integration
+For HPC environments (like Leonardo Booster), running OLMES directly from the trainer process can cause GPU Out-of-Memory (OOM) errors. You can decouple evaluation by creating a wrapper script (`scripts/evaluation/run_olmes_sbatch.sh`) that submits an independent SLURM job using `sbatch --wait`. Because `DownstreamEvaluator` uses `subprocess.Popen` asynchronously, the wrapper script will wait in the background on the training node without blocking the training loop!
+
+> [!IMPORTANT]
+> **Nested SLURM Job Environment Isolation (`--export=NONE`)**
+> When submitting the nested evaluation job using `sbatch` from within a running SLURM training job, the nested job inherits the parent job's environment variables (such as CUDA variables, `RANK`, `WORLD_SIZE`, `MASTER_ADDR`, etc.) by default. This will cause the evaluation job to fail or behave incorrectly.
+>
+> To prevent environment leakage, you **must** include `#SBATCH --export=NONE` in the nested `sbatch` script header. This ensures the evaluation job starts with a clean, isolated environment.
+
+### YAML Configuration
+```yaml
+downstream_evaluator:
+  component_key: downstream_evaluator
+  variant_key: default
+  config:
+    tokenizer:
+      instance_key: tokenizer
+      pass_type: BY_REFERENCE
+    tasks:
+      - "arc_challenge::olmes"
+      - "hellaswag::olmes"
+    eval_interval: 100
+    checkpoint_dir: ${settings.paths.experiments_root_path}/${settings.experiment_id}
+    global_rank: ${settings.cuda_env.global_rank}
+    olmes_command_template: "bash scripts/evaluation/run_olmes_sbatch.sh {hf_model_dir} '{tasks}' {step} 1024 1"
+```
+
+---
+
+## System Integration Summary
+
+For context on how these components are wired into the system, the following files handle the integration:
+
+1. **`src/modalities/trainer.py`**
+   - `conversion_callback` was added to `train()` signature.
+   - Pre-loop and in-loop execution order was explicitly set to: `checkpointing_callback` -> `conversion_callback` -> `downstream_evaluation_callback`.
+
+2. **`src/modalities/gym.py`**
+   - Threads `conversion_callback` through `Gym.run()` and passes it down to `self.trainer.train()`.
+
+3. **`src/modalities/main.py` (Lines 227-249)**
+   - Resolves `components.model_converter.convert` and `components.downstream_evaluator.evaluate`.
+   - Passes them into `gym.run()`.
+   - **Post-Training Wait** (Lines 244-249): At the very end of `run()`, explicitly calls `components.downstream_evaluator.wait_for_evaluations()` with prominent `print_rank_0` logging to ensure training does not exit until evaluations complete.
+
+4. **`src/modalities/config/config.py`**
+   - Defines Pydantic models `ModelConverterConfig` and `DownstreamEvaluatorConfig`.
+
+5. **`src/modalities/config/instantiation_models.py`**
+   - Adds `model_converter` and `downstream_evaluator` fields to `TrainingComponentsInstantiationModel`.
+
+6. **`src/modalities/registry/components.py`**
+   - Registers both classes to the `"default"` component registry.
+
+7. **`src/modalities/conversion/gpt2/convert_gpt2.py` (Lines 105-112)**
+   - Updated to support Hugging Face tokenizers (`pretrained_hf_tokenizer`) alongside SentencePiece. Detects tokenizer configs and saves `vocab.json` / `tokenizer.json` directly to the `hf_checkpoint` directory.
+
+8. **`tests/test_downstream_evaluator.py`**
+   - Contains comprehensive tests mocking the `subprocess` calls and verifying interval gating, rank gating, and directory existence logic.
+
+---
+
+## 3. Precaching Datasets (Offline Environments)
+
+If your compute cluster nodes do not have internet access, you must precache the Hugging Face datasets that OLMES requires. We provide a generalized script `scripts/evaluation/precache_tasks.py` that you can run on a login node (or any environment with internet access).
+
+### Usage
+
+Activate your evaluation environment (virtualenv, conda, or your Singularity container shell) and set the `HF_DATASETS_CACHE` and `HF_HOME` variables to a location accessible by your compute nodes.
+
+```bash
+# 1. Activate your python environment (e.g. venv where olmes is installed)
+source /path/to/olmes/venv/bin/activate
+export PYTHONPATH=/path/to/olmes/venv/lib/python3.12/site-packages:$PYTHONPATH
+
+# 2. Point Hugging Face to a shared scratch space or cache directory
+export HF_DATASETS_CACHE="/path/to/shared/hf_cache"
+export HF_HOME="/path/to/shared/hf_cache"
+export HF_TOKEN="your_hf_access_token"  # If needed for gated models/datasets
+
+# 3. Define the tasks you need
+export OLMES_TASKS="arc_challenge:rc::olmes:full hellaswag:rc::olmes:full gsm8k::olmes"
+
+# 4. Run the precache script
+python scripts/evaluation/precache_tasks.py --tasks $OLMES_TASKS
+```
+
+This script will resolve the tasks via OLMES and download all required datasets to your cache directory. When you run your training job via `sbatch`, ensure the compute nodes also set `HF_DATASETS_CACHE` and `HF_HOME` to the exact same shared directory.
diff --git a/scripts/evaluation/precache_tasks.py b/scripts/evaluation/precache_tasks.py
new file mode 100644
index 000000000..51949e8ae
--- /dev/null
+++ b/scripts/evaluation/precache_tasks.py
@@ -0,0 +1,62 @@
+import argparse
+import copy
+import os
+
+from datasets import load_dataset
+from oe_eval.configs.tasks import TASK_CONFIGS
+from oe_eval.launch import resolve_task_suite
+from oe_eval.run_eval import load_task
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Precache OLMES tasks and required HF datasets.")
+    parser.add_argument(
+        "--tasks", 
+        nargs="+", 
+        required=True, 
+        help="List of OLMES tasks to precache (e.g. arc_challenge:rc::olmes:full hellaswag:rc::olmes:full)"
+    )
+    args = parser.parse_args()
+
+    hf_home = os.environ.get("HF_DATASETS_CACHE", os.environ.get("HF_HOME", "~/.cache/huggingface"))
+    print(f"HF_DATASETS_CACHE is set to: {hf_home}")
+
+    # ---- Part 1: OLMES tasks ----
+    print("\n--- Caching OLMES tasks ---")
+    all_tasks = []
+    for t in args.tasks:
+        try:
+            all_tasks += resolve_task_suite(t, {})
+        except Exception as e:
+            print(f"!! could not resolve {t}: {e}")
+
+    print(f"\nWill download {len(all_tasks)} tasks to {hf_home}")
+    for task_name in all_tasks:
+        if task_name not in TASK_CONFIGS:
+            print(f"?? not in TASK_CONFIGS: {task_name}")
+            continue
+        try:
+            cfg = copy.deepcopy(TASK_CONFIGS[task_name])
+            task = load_task(cfg, ".")
+            print(f"-> downloading {task_name}")
+            task.download()
+            print(f"   done")
+        except Exception as e:
+            print(f"!! failed {task_name}: {e}")
+
+    # ---- Part 2: pseudo-sources used by paloma_diagnostics.py ----
+    print("\n--- Caching diagnostics pseudo-source datasets (gsm8k, trivia_qa) ---")
+    specs = [
+        ('gsm8k',     'main',         'test'),
+        ('trivia_qa', 'rc.nocontext', 'validation'),
+    ]
+    for path, config, split in specs:
+        try:
+            print(f"-> downloading {path} ({config}, {split})")
+            load_dataset(path, config, split=split)
+            print(f"   done")
+        except Exception as e:
+            print(f"!! failed {path}: {e}")
+
+if __name__ == "__main__":
+    main()
diff --git a/src/modalities/batch.py b/src/modalities/batch.py
index bb55f245a..834ad6b1e 100644
--- a/src/modalities/batch.py
+++ b/src/modalities/batch.py
@@ -54,6 +54,26 @@ def __len__(self) -> int:
         return self.samples[key].shape[self.batch_dim]
 
 
+def _apply_to(val, device):
+    if isinstance(val, torch.Tensor):
+        return val.to(device)
+    elif isinstance(val, dict):
+        return {k: _apply_to(v, device) for k, v in val.items()}
+    elif isinstance(val, list):
+        return [_apply_to(v, device) for v in val]
+    return val
+
+
+def _apply_detach(val):
+    if isinstance(val, torch.Tensor):
+        return val.detach()
+    elif isinstance(val, dict):
+        return {k: _apply_detach(v) for k, v in val.items()}
+    elif isinstance(val, list):
+        return [_apply_detach(v) for v in val]
+    return val
+
+
 @dataclass
 class InferenceResultBatch(Batch, TorchDeviceMixin):
     """Stores targets and predictions of an entire batch."""
@@ -71,12 +91,12 @@ def device(self) -> torch.device:
         return self.targets[key].device
 
     def to(self, device: torch.device):
-        self.predictions = {k: v.to(device) for k, v in self.predictions.items()}
-        self.targets = {k: v.to(device) for k, v in self.targets.items()}
+        self.predictions = {k: _apply_to(v, device) for k, v in self.predictions.items()}
+        self.targets = {k: _apply_to(v, device) for k, v in self.targets.items()}
 
     def detach(self):
-        self.targets = {k: v.detach() for k, v in self.targets.items()}
-        self.predictions = {k: v.detach() for k, v in self.predictions.items()}
+        self.targets = {k: _apply_detach(v) for k, v in self.targets.items()}
+        self.predictions = {k: _apply_detach(v) for k, v in self.predictions.items()}
 
     def get_predictions(self, key: str) -> torch.Tensor:
         if key not in self.predictions:
@@ -89,8 +109,13 @@ def get_targets(self, key: str) -> torch.Tensor:
         return self.targets[key]
 
     def __len__(self) -> int:
-        key = list(self.predictions.keys())[0]
-        return self.predictions[key].shape[self.batch_dim]
+        for v in self.predictions.values():
+            if isinstance(v, torch.Tensor):
+                return v.shape[self.batch_dim]
+        for v in self.targets.values():
+            if isinstance(v, torch.Tensor):
+                return v.shape[self.batch_dim]
+        raise ValueError("No tensor found in predictions or targets to determine batch length")
 
 
 @dataclass
diff --git a/src/modalities/config/config.py b/src/modalities/config/config.py
index 42a19b99a..cc623e336 100644
--- a/src/modalities/config/config.py
+++ b/src/modalities/config/config.py
@@ -520,6 +520,22 @@ class ParallelDegreeConfig(BaseModel):
     parallelism_methods: list[ParallelismDegrees]
 
 
+class ModelConverterConfig(BaseModel):
+    command_template: str
+    checkpoint_dir: Path
+    global_rank: Annotated[int, Field(strict=True, ge=0)]
+    eval_interval: Annotated[int, Field(strict=True, gt=0)]
+
+
+class DownstreamEvaluatorConfig(BaseModel):
+    tokenizer: PydanticTokenizerIFType
+    tasks: list[str]
+    eval_interval: Annotated[int, Field(strict=True, gt=0)]
+    checkpoint_dir: Path
+    global_rank: Annotated[int, Field(strict=True, ge=0)]
+    olmes_command_template: str
+
+
 # Recursive type representing arbitrary-depth YAML config structures.
 YAMLPrimitive = str | int | float | bool | None
 YAMLValue: TypeAlias = YAMLPrimitive | Path | list["YAMLValue"] | dict[str, "YAMLValue"]
diff --git a/src/modalities/config/instantiation_models.py b/src/modalities/config/instantiation_models.py
index fd7fd3b78..107394349 100644
--- a/src/modalities/config/instantiation_models.py
+++ b/src/modalities/config/instantiation_models.py
@@ -22,6 +22,8 @@
     PydanticSteppableProfilerIFType,
     PydanticTextInferenceComponentType,
     PydanticTokenizerIFType,
+    PydanticDownstreamEvaluatorType,
+    PydanticModelConverterType,
 )
 from modalities.config.utils import parse_torch_device
 from modalities.dataloader.dataset import Dataset
@@ -192,6 +194,8 @@ def _check_last_step_checkpointed(self) -> "TrainingComponentsInstantiationModel
     mfu_calculator: PydanticMFUCalculatorABCType | None = None
     scheduled_pipeline: PydanticPipelineType | None = None
     device_mesh: PydanticDeviceMeshIFType | None = None
+    downstream_evaluator: Optional[PydanticDownstreamEvaluatorType] = None
+    model_converter: Optional[PydanticModelConverterType] = None
     model_raw: PydanticPytorchModuleType
 
     @model_validator(mode="after")
diff --git a/src/modalities/config/pydantic_if_types.py b/src/modalities/config/pydantic_if_types.py
index 90b7ca951..217c2f783 100644
--- a/src/modalities/config/pydantic_if_types.py
+++ b/src/modalities/config/pydantic_if_types.py
@@ -19,6 +19,8 @@
 from modalities.checkpointing.stateful.app_state import AppState
 from modalities.dataloader.collate_fns.collate_if import CollateFnIF
 from modalities.dataloader.dataloader import LLMDataLoader
+from modalities.conversion.model_converter import ModelConverter
+from modalities.evaluator import DownstreamEvaluator
 from modalities.inference.text.inference_component import TextInferenceComponent
 from modalities.logging_broker.subscriber import MessageSubscriberIF
 from modalities.loss_functions import Loss
@@ -98,3 +100,5 @@ def __get_pydantic_core_schema__(
     torch.utils.hooks.RemovableHandle, PydanticThirdPartyTypeIF(torch.utils.hooks.RemovableHandle)
 ]
 PydanticDebuggingType = Annotated[Debugging, PydanticThirdPartyTypeIF(Debugging)]
+PydanticDownstreamEvaluatorType = Annotated[DownstreamEvaluator, PydanticThirdPartyTypeIF(DownstreamEvaluator)]
+PydanticModelConverterType = Annotated[ModelConverter, PydanticThirdPartyTypeIF(ModelConverter)]
diff --git a/src/modalities/conversion/gpt2/configuration_gpt2.py b/src/modalities/conversion/gpt2/configuration_gpt2.py
index 7663cd227..d5ecbc8bc 100644
--- a/src/modalities/conversion/gpt2/configuration_gpt2.py
+++ b/src/modalities/conversion/gpt2/configuration_gpt2.py
@@ -181,10 +181,12 @@ def __init__(
         attention_dropout=0.0,
         mlp_bias=False,
         head_dim=None,
+        norm_type="layer_norm",
+        use_qk_norm=False,
+        qk_norm_dim=None,
         **kwargs,
     ):
-        if rms_norm_eps is not None:
-            raise ValueError("RMSNorm is not supported in GPT2 model.")
+        self.norm_type = norm_type
         self.vocab_size = vocab_size
         self.max_position_embeddings = max_position_embeddings
         self.hidden_size = hidden_size
@@ -211,6 +213,8 @@ def __init__(
         self.attention_dropout = attention_dropout
         self.mlp_bias = mlp_bias
         self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads
+        self.use_qk_norm = use_qk_norm
+        self.qk_norm_dim = qk_norm_dim
         # Validate the correctness of rotary position embeddings parameters
         # BC: if there is a 'type' field, copy it it to 'rope_type'.
         if self.rope_scaling is not None and "type" in self.rope_scaling:
diff --git a/src/modalities/conversion/gpt2/conversion_model.py b/src/modalities/conversion/gpt2/conversion_model.py
index f44ff33e6..f5999faf4 100644
--- a/src/modalities/conversion/gpt2/conversion_model.py
+++ b/src/modalities/conversion/gpt2/conversion_model.py
@@ -23,7 +23,43 @@ def convert_model_checkpoint(modalities_config: dict) -> tuple[GPT2ForCausalLM,
     """
     gpt2_config = convert_model_config(modalities_config)
     hf_model = GPT2ForCausalLM(gpt2_config).to(dtype=torch.bfloat16)
-    modalities_model = get_model_from_config(modalities_config, model_type=ModelTypeEnum.CHECKPOINTED_MODEL)
+    model_config = modalities_config["model_raw" if "model_raw" in modalities_config else "model"]
+    checkpoint_path = None
+    if "checkpointed_model" in modalities_config:
+        checkpoint_path = modalities_config["checkpointed_model"].get("config", {}).get("checkpoint_path")
+
+    if checkpoint_path and not ("variant_key" in modalities_config.get("checkpointed_model", {})):
+        # Load state dict manually if variant_key is missing
+        if "model" not in modalities_config and "model_raw" in modalities_config:
+            modalities_config["model"] = modalities_config["model_raw"]
+        modalities_model = get_model_from_config(modalities_config, model_type=ModelTypeEnum.MODEL)
+        from pathlib import Path
+        if Path(checkpoint_path).is_dir():
+            from torch.distributed.checkpoint.default_planner import _EmptyStateDictLoadPlanner
+            from torch.distributed.checkpoint.filesystem import FileSystemReader
+            from torch.distributed.checkpoint.state_dict_loader import _load_state_dict
+            sd = {}
+            planner = _EmptyStateDictLoadPlanner(keys=["app.model"], allow_partial_load=True)
+            _load_state_dict(sd, storage_reader=FileSystemReader(checkpoint_path), planner=planner, no_dist=True)
+            model_sd = sd.get("app", {}).get("model", sd)
+        else:
+            ckpt = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
+            model_sd = ckpt
+            for key in ("model_state_dict", "state_dict", "model"):
+                if key in ckpt and isinstance(ckpt[key], dict):
+                    model_sd = ckpt[key]
+                    break
+        
+        out = {}
+        for k, v in model_sd.items():
+            if k.startswith("module."):
+                k = k[len("module."):]
+            out[k] = v
+        missing, unexpected = modalities_model.load_state_dict(out, strict=False)
+        print("Missing keys:", missing)
+        print("Unexpected keys:", unexpected)
+    else:
+        modalities_model = get_model_from_config(modalities_config, model_type=ModelTypeEnum.CHECKPOINTED_MODEL)
     _copy_weights_model(hf_model, modalities_model)
     return hf_model, modalities_model
 
@@ -43,6 +79,14 @@ def convert_model_config(modalities_config: dict) -> GPT2Config:
     _check_conversion_criteria(config)
 
     ffn_norm_key = "ffn_norm_config"
+    norm_type = config[ffn_norm_key].get("norm_type", "layer_norm")
+
+    qk_norm_cfg = config.get("attention_config", {}).get("qk_norm_config")
+    use_qk_norm = qk_norm_cfg is not None
+    qk_norm_dim = None
+    if use_qk_norm:
+        qk_cfg = qk_norm_cfg.get("config", {})
+        qk_norm_dim = qk_cfg.get("ndim", qk_cfg.get("normalized_shape"))
 
     return GPT2Config(
         vocab_size=config["vocab_size"],
@@ -57,12 +101,15 @@ def convert_model_config(modalities_config: dict) -> GPT2Config:
         attention_bias=config["bias"],
         mlp_bias=config["bias"],
         hidden_act="silu",
+        norm_type=norm_type,
         layer_norm_eps=_get_layer_norm_value(config[ffn_norm_key]["config"], "eps"),
         layer_norm_elementwise_affine=_get_layer_norm_value(config[ffn_norm_key]["config"], "elementwise_affine"),
         layer_norm_bias=_get_layer_norm_value(config[ffn_norm_key]["config"], "bias"),
         max_position_embeddings=config["sequence_length"],
         rope_theta=config["attention_config"]["qkv_transforms"][0]["config"]["base_freq"],
         _attn_implementation=_map_attention_type(config),
+        use_qk_norm=use_qk_norm,
+        qk_norm_dim=qk_norm_dim,
         output_attentions=False,
     )
 
@@ -80,6 +127,7 @@ def check_converted_model(hf_model: GPT2ForCausalLM, modalities_model: GPT2LLM,
         input_ids = torch.randint(0, vocab_size, (1, modalities_model.sequence_length), device=hf_model.device)
         inputs = {modalities_model.sample_key: input_ids.to(modalities_model.transformer.wte.weight.device)}
 
+        modalities_model.to(dtype=hf_model.dtype, device=hf_model.device)
         with torch.no_grad():
             llama_logits = hf_model(input_ids=input_ids).logits.to("cpu")
             modalities_logits = modalities_model(inputs)[modalities_model.prediction_key].to("cpu")
@@ -103,7 +151,7 @@ def _check_conversion_criteria(model_config: dict) -> None:
 
     norms = ["attention_norm_config", "ffn_norm_config", "lm_head_norm_config"]
     for norm in norms:
-        assert model_config[norm]["norm_type"] == "layer_norm"
+        assert model_config[norm]["norm_type"] in ["layer_norm", "rms_norm", "pytorch_rms_norm"]
 
     assert (
         len(set(_get_layer_norm_value(model_config[norm]["config"], "bias") for norm in norms)) == 1
@@ -122,13 +170,13 @@ def _get_layer_norm_value(config: dict, field: str) -> bool | float | int:
 
 
 def _map_attention_type(config: dict):
-    if config["attention_implementation"] == "pytorch_flash":
-        attention_impl = "sdpa"
-    elif config["attention_implementation"] == "manual":
-        attention_impl = "eager"
+    impl = config.get("attention_implementation", "default")
+    if impl in ("pytorch_flash", "default"):
+        return "sdpa"
+    elif impl == "manual":
+        return "eager"
     else:
-        raise ValueError(f"Unknown or unsupported attention implementation {config['attention_implementation']}.")
-    return attention_impl
+        raise ValueError(f"Unknown attention_implementation: {impl}")
 
 
 def _copy_weights_model(hf_model: GPT2ForCausalLM, modalities_model: GPT2LLM):
@@ -164,11 +212,16 @@ def _copy_weights_mlp(hf_layer: GPT2DecoderLayer, modalities_layer: GPT2Block):
 def _copy_weights_layer_norms(hf_layer: GPT2DecoderLayer, modalities_layer: GPT2Block):
     _copy_weights_base_modules(hf_layer.input_layernorm, modalities_layer.attention_norm)
     _copy_weights_base_modules(hf_layer.post_attention_layernorm, modalities_layer.ffn_norm)
+    if getattr(hf_layer.self_attn, "q_norm", None) is not None:
+        _copy_weights_base_modules(hf_layer.self_attn.q_norm, modalities_layer.attn.q_norm)
+        _copy_weights_base_modules(hf_layer.self_attn.k_norm, modalities_layer.attn.k_norm)
 
 
-def _copy_weights_base_modules(m1: nn.Linear | nn.LayerNorm, m2: nn.Linear | nn.LayerNorm):
+def _copy_weights_base_modules(m1: nn.Linear | nn.LayerNorm | nn.Module, m2: nn.Linear | nn.LayerNorm | nn.Module):
     assert m1.weight.shape == m2.weight.shape
-    assert (m1.bias is None and m2.bias is None) or m1.bias.shape == m2.bias.shape
+    m1_bias = getattr(m1, "bias", None)
+    m2_bias = getattr(m2, "bias", None)
+    assert (m1_bias is None and m2_bias is None) or m1_bias.shape == m2_bias.shape
     m1.weight.data.copy_(m2.weight.data)
-    if m1.bias is not None:
-        m1.bias.data.copy_(m2.bias.data)
+    if m1_bias is not None:
+        m1_bias.data.copy_(m2_bias.data)
diff --git a/src/modalities/conversion/gpt2/convert_gpt2.py b/src/modalities/conversion/gpt2/convert_gpt2.py
index d16f55ce9..649b91ac7 100644
--- a/src/modalities/conversion/gpt2/convert_gpt2.py
+++ b/src/modalities/conversion/gpt2/convert_gpt2.py
@@ -38,6 +38,7 @@ def convert_gpt2(
     num_testruns: int = 0,
     device_modalities: str = "cpu",
     device_hf: str = "cpu",
+    checkpoint_path: str | None = None,
 ) -> None:
     """Takes a modalities gpt2 model and converts it to a Huggingface transformers model.
        The provided config yaml file should contain the model_raw or model section with the model configuration.
@@ -56,6 +57,11 @@ def convert_gpt2(
         modalities_config = load_app_config_dict(
             Path(modalities_config_path), experiment_id="-1", experiments_root_path=Path(tmpdir)
         )
+        if checkpoint_path is not None:
+            if "checkpointed_model" not in modalities_config:
+                modalities_config["checkpointed_model"] = {"config": {}}
+            modalities_config["checkpointed_model"]["config"]["checkpoint_path"] = checkpoint_path
+
         hf_model, modalities_model = convert_model_checkpoint(modalities_config)
 
     if num_testruns > 0:
@@ -69,12 +75,20 @@ def convert_gpt2(
     sentence_piece_tokenizer_configs = {
         key: subconfig
         for key, subconfig in modalities_config.items()
-        if "component_key" in subconfig
+        if isinstance(subconfig, dict) and "component_key" in subconfig
         and subconfig["component_key"] == "tokenizer"
         and subconfig["variant_key"] == "pretrained_sp_tokenizer"
     }
 
-    if len(sentence_piece_tokenizer_configs) > 1:
+    hf_tokenizer_configs = {
+        key: subconfig
+        for key, subconfig in modalities_config.items()
+        if isinstance(subconfig, dict) and "component_key" in subconfig
+        and subconfig["component_key"] == "tokenizer"
+        and subconfig["variant_key"] == "pretrained_hf_tokenizer"
+    }
+
+    if len(sentence_piece_tokenizer_configs) + len(hf_tokenizer_configs) > 1:
         raise ValueError(
             "Multiple tokenizer configs found. Please specify only one tokenizer config in the modalities config file."
         )
@@ -88,6 +102,14 @@ def convert_gpt2(
         hf_model.config.bos_token_id = bos_token_id
         hf_model.config.eos_token_id = eos_token_id
         hf_model.config.pad_token_id = pad_token_id
+    elif len(hf_tokenizer_configs) == 1:
+        from transformers import AutoTokenizer
+        tokenizer_name = modalities_config["tokenizer"]["config"]["pretrained_model_name_or_path"]
+        tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+        tokenizer.save_pretrained(output_dir)
+        hf_model.config.bos_token_id = tokenizer.bos_token_id
+        hf_model.config.eos_token_id = tokenizer.eos_token_id
+        hf_model.config.pad_token_id = tokenizer.pad_token_id
     else:
         logger.warning("No tokenizer specified in the config. Skipping tokenizer conversion.")
     hf_model.config.auto_map = {
@@ -110,6 +132,7 @@ def convert_gpt2(
     parser.add_argument("--num_testruns", type=int, default=0, help="Number of test runs to perform.")
     parser.add_argument("--device_modalities", type=str, default="cpu", help="Device for the modalities model.")
     parser.add_argument("--device_hf", type=str, default="cpu", help="Device for the Hugging Face model.")
+    parser.add_argument("--checkpoint_path", type=str, default=None, help="Path to the model checkpoint. Overrides config.")
 
     args = parser.parse_args()
 
@@ -119,4 +142,5 @@ def convert_gpt2(
         args.num_testruns,
         args.device_modalities,
         args.device_hf,
+        args.checkpoint_path,
     )
diff --git a/src/modalities/conversion/gpt2/modeling_gpt2.py b/src/modalities/conversion/gpt2/modeling_gpt2.py
index f6aa77ab1..98d4141d7 100644
--- a/src/modalities/conversion/gpt2/modeling_gpt2.py
+++ b/src/modalities/conversion/gpt2/modeling_gpt2.py
@@ -88,6 +88,23 @@ def forward(self, x, position_ids):
         return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
 
 
+class GPT2RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        GPT2RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
 def rotate_half(x):
     """Rotates half the hidden dims of the input."""
     x1 = x[..., : x.shape[-1] // 2]
@@ -203,6 +220,17 @@ def __init__(self, config: GPT2Config, layer_idx: int):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
+        if getattr(config, "use_qk_norm", False):
+            qk_norm_dim = getattr(config, "qk_norm_dim", None) or self.head_dim
+            if getattr(config, "norm_type", "layer_norm") == "pytorch_rms_norm":
+                self.q_norm = nn.RMSNorm(qk_norm_dim, eps=config.layer_norm_eps)
+                self.k_norm = nn.RMSNorm(qk_norm_dim, eps=config.layer_norm_eps)
+            else:
+                self.q_norm = GPT2RMSNorm(qk_norm_dim, eps=config.layer_norm_eps)
+                self.k_norm = GPT2RMSNorm(qk_norm_dim, eps=config.layer_norm_eps)
+        else:
+            self.q_norm = None
+            self.k_norm = None
 
     def forward(
         self,
@@ -223,6 +251,10 @@ def forward(
         cos, sin = position_embeddings
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
+        if getattr(self, "q_norm", None) is not None:
+            query_states = self.q_norm(query_states)
+            key_states = self.k_norm(key_states)
+
         if past_key_value is not None:
             # sin and cos are specific to RoPE models; cache_position needed for the static cache
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
@@ -256,18 +288,25 @@ def __init__(self, config: GPT2Config, layer_idx: int):
         self.self_attn = LlamaAttention(config=config, layer_idx=layer_idx)
 
         self.mlp = LlamaMLP(config)
-        self.input_layernorm = nn.LayerNorm(
-            config.hidden_size,
-            eps=config.layer_norm_eps,
-            elementwise_affine=config.layer_norm_elementwise_affine,
-            bias=config.layer_norm_bias,
-        )
-        self.post_attention_layernorm = nn.LayerNorm(
-            config.hidden_size,
-            eps=config.layer_norm_eps,
-            elementwise_affine=config.layer_norm_elementwise_affine,
-            bias=config.layer_norm_bias,
-        )
+        if getattr(config, "norm_type", "layer_norm") == "pytorch_rms_norm":
+            self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
+            self.post_attention_layernorm = nn.RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
+        elif getattr(config, "norm_type", "layer_norm") == "rms_norm":
+            self.input_layernorm = GPT2RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
+            self.post_attention_layernorm = GPT2RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
+        else:
+            self.input_layernorm = nn.LayerNorm(
+                config.hidden_size,
+                eps=config.layer_norm_eps,
+                elementwise_affine=config.layer_norm_elementwise_affine,
+                bias=config.layer_norm_bias,
+            )
+            self.post_attention_layernorm = nn.LayerNorm(
+                config.hidden_size,
+                eps=config.layer_norm_eps,
+                elementwise_affine=config.layer_norm_elementwise_affine,
+                bias=config.layer_norm_bias,
+            )
 
     def forward(
         self,
@@ -333,19 +372,23 @@ def __init__(self, config: GPT2Config):
         self.layers = nn.ModuleList(
             [GPT2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
         )
-        self.norm = nn.LayerNorm(
-            config.hidden_size,
-            eps=config.layer_norm_eps,
-            elementwise_affine=config.layer_norm_elementwise_affine,
-            bias=config.layer_norm_bias,
-        )
+        if getattr(config, "norm_type", "layer_norm") == "pytorch_rms_norm":
+            self.norm = nn.RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
+        elif getattr(config, "norm_type", "layer_norm") == "rms_norm":
+            self.norm = GPT2RMSNorm(config.hidden_size, eps=config.layer_norm_eps)
+        else:
+            self.norm = nn.LayerNorm(
+                config.hidden_size,
+                eps=config.layer_norm_eps,
+                elementwise_affine=config.layer_norm_elementwise_affine,
+                bias=config.layer_norm_bias,
+            )
         self.rotary_emb = LlamaRotaryEmbedding(config=config)
         self.gradient_checkpointing = False
 
         # Initialize weights and apply final processing
         self.post_init()
 
-    @check_model_inputs
     @auto_docstring
     def forward(
         self,
diff --git a/src/modalities/conversion/model_converter.py b/src/modalities/conversion/model_converter.py
new file mode 100644
index 000000000..377d22292
--- /dev/null
+++ b/src/modalities/conversion/model_converter.py
@@ -0,0 +1,95 @@
+import json
+import logging
+import subprocess
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+
+class ModelConverter:
+    """Converts a Modalities checkpoint to HF format by running a user-provided command template.
+
+    The command_template is a shell command string with placeholders:
+        {checkpoint_path} - resolved from last_checkpoint_info.json
+        {output_dir}      - <checkpoint_path>/hf_checkpoint
+        {modalities_config} - path to the modalities YAML config (if available)
+
+    Example command_templates:
+        "python convert_gpt2.py {modalities_config} {output_dir}"
+        "python convert_adaptive_gpt.py {checkpoint_path} {output_dir} --modalities_config {modalities_config}"
+    """
+
+    def __init__(
+        self,
+        command_template: str,
+        checkpoint_dir: Path,
+        global_rank: int,
+        eval_interval: int,
+    ) -> None:
+        self.command_template = command_template
+        self.checkpoint_dir = Path(checkpoint_dir)
+        self.global_rank = global_rank
+        self.eval_interval = eval_interval
+
+    def convert(self, num_train_steps_done: int) -> None:
+        """Run the conversion command if the current step matches the eval interval.
+
+        Args:
+            num_train_steps_done: Number of training steps completed so far.
+        """
+        if num_train_steps_done == 0 or num_train_steps_done % self.eval_interval != 0:
+            return
+        if self.global_rank != 0:
+            return
+
+        checkpoint_path = self._get_latest_checkpoint_path()
+        if checkpoint_path is None:
+            logger.warning("No checkpoint info found, skipping conversion.")
+            return
+
+        output_dir = checkpoint_path / "hf_checkpoint"
+        if output_dir.exists():
+            logger.info(f"HF checkpoint already exists at {output_dir}, skipping conversion.")
+            return
+
+        cmd = self.command_template.format(
+            checkpoint_path=str(checkpoint_path),
+            output_dir=str(output_dir),
+            modalities_config=str(self._find_config_in_checkpoint(checkpoint_path)),
+        )
+
+        logger.info(f"Running model conversion: {cmd}")
+        try:
+            subprocess.run(cmd, shell=True, check=True)
+            logger.info(f"Conversion complete. HF checkpoint saved to {output_dir}")
+        except subprocess.CalledProcessError as e:
+            logger.error(f"Model conversion failed with return code {e.returncode}: {e}")
+
+    def _get_latest_checkpoint_path(self) -> Path | None:
+        """Read last_checkpoint_info.json to find the latest checkpoint path."""
+        info_file = self.checkpoint_dir / "last_checkpoint_info.json"
+        if not info_file.exists():
+            return None
+
+        with open(info_file, "r", encoding="utf-8") as f:
+            info = json.load(f)
+
+        # DCP checkpoints use "checkpoint_folder_path", FSDP1 uses "model_checkpoint_path"
+        checkpoint_path_str = info.get("checkpoint_folder_path") or info.get("model_checkpoint_path")
+        if checkpoint_path_str is None:
+            return None
+
+        path = Path(checkpoint_path_str)
+        # For FSDP1, model_checkpoint_path points to the .bin file; we want the parent directory
+        if path.is_file():
+            path = path.parent
+        return path
+
+    @staticmethod
+    def _find_config_in_checkpoint(checkpoint_path: Path) -> Path | None:
+        """Look for a YAML config file inside or next to the checkpoint directory."""
+        for search_dir in [checkpoint_path, checkpoint_path.parent]:
+            for f in search_dir.iterdir():
+                if f.suffix in (".yaml", ".yml") and not f.name.endswith(".resolved"):
+                    return f
+        return Path("")
diff --git a/src/modalities/evaluator.py b/src/modalities/evaluator.py
index fb9bdc0d3..4fc7ee327 100644
--- a/src/modalities/evaluator.py
+++ b/src/modalities/evaluator.py
@@ -1,8 +1,14 @@
+import json
+import logging
+import subprocess
+from pathlib import Path
 from typing import Callable
 
 import torch
 import torch.distributed as dist
 import torch.nn as nn
+
+from modalities.tokenization.tokenizer_wrapper import TokenizerWrapper
 from torch.distributed.device_mesh import DeviceMesh
 
 from modalities.batch import DatasetBatch, EvaluationResultBatch, InferenceResultBatch, ResultItem
@@ -15,6 +21,8 @@
 from modalities.running_env.fsdp.reducer import Reducer
 from modalities.util import TimeRecorder
 
+logger = logging.getLogger(__name__)
+
 
 class Evaluator:
     """Evaluator class which is responsible for evaluating the model on a set of datasets"""
@@ -197,3 +205,135 @@ def _publish_evaluation_result(
         evaluation_result_publisher.publish_message(
             payload=evaluation_result, message_type=MessageTypes.EVALUATION_RESULT
         )
+
+
+class DownstreamEvaluator:
+    """Evaluator that runs OLMES on HF checkpoints produced by the conversion callback.
+
+    Checks if an ``hf_checkpoint`` folder exists inside the latest checkpoint directory
+    (as written by ``ModelConverter``).  If it does, the configured OLMES command template
+    is executed via subprocess.
+    """
+
+    def __init__(
+        self,
+        tokenizer: TokenizerWrapper,
+        tasks: list[str],
+        eval_interval: int,
+        checkpoint_dir: Path,
+        global_rank: int,
+        olmes_command_template: str,
+    ) -> None:
+        self.tokenizer = tokenizer
+        self.tasks = tasks
+        self.eval_interval = eval_interval
+        self.checkpoint_dir = Path(checkpoint_dir)
+        self.global_rank = global_rank
+        self.olmes_command_template = olmes_command_template
+        self.active_processes: list[tuple[subprocess.Popen, int, Path]] = []
+
+    def evaluate(self, num_train_steps_done: int) -> None:
+        if num_train_steps_done == 0 or num_train_steps_done % self.eval_interval != 0:
+            return
+        if self.global_rank != 0:
+            return
+
+        hf_model_dir = self._find_hf_checkpoint()
+        if hf_model_dir is None:
+            logger.warning(
+                f"No hf_checkpoint found in {self.checkpoint_dir} at step {num_train_steps_done}, "
+                "skipping downstream evaluation."
+            )
+            return
+
+        tasks_str = " ".join(self.tasks)
+        cmd = self.olmes_command_template.format(
+            hf_model_dir=str(hf_model_dir),
+            tasks=tasks_str,
+            step=num_train_steps_done,
+        )
+
+        logger.info(f"Running downstream evaluation: {cmd}")
+        try:
+            p = subprocess.Popen(cmd, shell=True)
+            self.active_processes.append((p, num_train_steps_done, hf_model_dir))
+            logger.info(f"Downstream evaluation launched for step {num_train_steps_done}.")
+        except Exception as e:
+            logger.error(f"Failed to launch downstream evaluation: {e}")
+
+    def wait_for_evaluations(self) -> None:
+        if not hasattr(self, "active_processes") or not self.active_processes:
+            return
+
+        logger.info(f"Waiting for {len(self.active_processes)} downstream evaluations to finish...")
+        for p, step, hf_model_dir in self.active_processes:
+            p.wait()
+            if p.returncode == 0:
+                self._sync_metrics_to_wandb(step, hf_model_dir)
+            else:
+                logger.warning(f"Downstream evaluation for step {step} exited with code {p.returncode}, skipping W&B sync.")
+        logger.info("All downstream evaluations finished.")
+        self.active_processes = []
+
+    def _sync_metrics_to_wandb(self, step: int, hf_model_dir: Path) -> None:
+        """Parse OLMES metrics-all.jsonl and log primary scores to the active W&B run."""
+        metrics_file = hf_model_dir / f"olmes_eval_{step}" / "metrics-all.jsonl"
+        if not metrics_file.exists():
+            logger.warning(f"No metrics file found at {metrics_file}, skipping W&B sync for step {step}.")
+            return
+
+        metrics_dict = {}
+        try:
+            with open(metrics_file, "r", encoding="utf-8") as f:
+                for line in f:
+                    obj = json.loads(line)
+                    alias = (
+                        obj.get("task_config", {}).get("metadata", {}).get("alias")
+                        or obj.get("task_name")
+                    )
+                    score = obj.get("metrics", {}).get("primary_score")
+                    if alias and score is not None:
+                        metrics_dict[f"downstream/{alias}"] = score
+        except Exception as e:
+            logger.error(f"Failed to parse metrics file {metrics_file}: {e}")
+            return
+
+        if not metrics_dict:
+            logger.warning(f"No metrics extracted from {metrics_file} for step {step}.")
+            return
+
+        try:
+            import wandb
+
+            if wandb.run is not None:
+                # Define a custom step metric so downstream/* metrics are decoupled from
+                # the global training step counter (which is already past these steps).
+                wandb.run.define_metric("downstream_step", hidden=True)
+                wandb.run.define_metric("downstream/*", step_metric="downstream_step")
+                metrics_dict["downstream_step"] = step
+                wandb.run.log(metrics_dict)
+                logger.info(f"Synced {len(metrics_dict)} OLMES metrics to W&B at step {step}: {metrics_dict}")
+            else:
+                logger.info(f"W&B not active, skipping metric sync for step {step}.")
+        except ImportError:
+            logger.info(f"wandb not installed, skipping metric sync for step {step}.")
+
+    def _find_hf_checkpoint(self) -> Path | None:
+        """Read last_checkpoint_info.json and check for hf_checkpoint subfolder."""
+        info_file = self.checkpoint_dir / "last_checkpoint_info.json"
+        if not info_file.exists():
+            return None
+
+        with open(info_file, "r", encoding="utf-8") as f:
+            info = json.load(f)
+
+        checkpoint_path_str = info.get("checkpoint_folder_path") or info.get("model_checkpoint_path")
+        if checkpoint_path_str is None:
+            return None
+
+        checkpoint_path = Path(checkpoint_path_str)
+        if checkpoint_path.is_file():
+            checkpoint_path = checkpoint_path.parent
+
+        hf_dir = checkpoint_path / "hf_checkpoint"
+        return hf_dir if hf_dir.exists() else None
diff --git a/src/modalities/gym.py b/src/modalities/gym.py
index 010c4ca60..d1835c6c8 100644
--- a/src/modalities/gym.py
+++ b/src/modalities/gym.py
@@ -1,6 +1,6 @@
 from datetime import datetime
 from functools import partial
-from typing import Callable
+from typing import Callable, Optional
 
 import torch.nn as nn
 
@@ -42,6 +42,8 @@ def run(
         evaluation_data_loaders: list[LLMDataLoader],
         checkpoint_saving: CheckpointSaving,
         scheduled_pipeline: Pipeline | None = None,
+        conversion_callback: Optional[Callable[[int], None]] = None,
+        downstream_evaluation_callback: Optional[Callable[[int], None]] = None,
     ):
         """Runs the model training, including evaluation and checkpointing.
 
@@ -55,6 +57,8 @@ def run(
             checkpoint_saving (CheckpointSaving): Routine for saving checkpoints.
             scheduled_pipeline (Pipeline | None, optional): In case of pipeline parallelism, this is used to
                 operate the model. Defaults to None.
+            conversion_callback (Optional[Callable[[int], None]]): A callback function for checkpoint-to-HF conversion.
+            downstream_evaluation_callback (Optional[Callable[[int], None]]): A callback function for downstream evaluation.
         """
         evaluation_callback: Callable[[int], None] = partial(
             self._run_evaluation,
@@ -80,6 +84,8 @@ def run(
             checkpointing_callback=checkpointing_callback,
             training_log_interval_in_steps=training_log_interval_in_steps,
             scheduled_pipeline=scheduled_pipeline,
+            conversion_callback=conversion_callback,
+            downstream_evaluation_callback=downstream_evaluation_callback,
         )
         print_rank_0(f"Training done at {datetime.now()}.")
 
diff --git a/src/modalities/logging_broker/subscriber_impl/results_subscriber.py b/src/modalities/logging_broker/subscriber_impl/results_subscriber.py
index d924a2e78..3f707c287 100644
--- a/src/modalities/logging_broker/subscriber_impl/results_subscriber.py
+++ b/src/modalities/logging_broker/subscriber_impl/results_subscriber.py
@@ -93,12 +93,18 @@ def consume_message(self, message: Message[EvaluationResultBatch]):
         """Consumes a message from a message broker."""
         eval_result = message.payload
 
+        def _wandb_key(tag: str, key: str) -> str:
+            """Route layer-specific metrics into a top-level layers/ panel."""
+            if "layer" in key:
+                return f"layers/{key}"
+            return f"{tag}/{key}"
+
         losses = {
-            f"{eval_result.dataloader_tag} {loss_key}": loss_values.value
+            _wandb_key(eval_result.dataloader_tag, loss_key): loss_values.value
             for loss_key, loss_values in eval_result.losses.items()
         }
         metrics = {
-            f"{eval_result.dataloader_tag} {metric_key}": metric_values.value
+            _wandb_key(eval_result.dataloader_tag, metric_key): metric_values.value
             for metric_key, metric_values in eval_result.metrics.items()
         }
         # TODO step is not semantically correct here. Need to check if we can rename step to num_samples
@@ -109,7 +115,7 @@ def consume_message(self, message: Message[EvaluationResultBatch]):
             data=metrics, step=eval_result.num_train_steps_done
         )  # (eval_result.train_local_sample_id + 1) * self.num_ranks)
         throughput_metrics = {
-            f"{eval_result.dataloader_tag} {metric_key}": metric_values.value
+            _wandb_key(eval_result.dataloader_tag, metric_key): metric_values.value
             for metric_key, metric_values in eval_result.throughput_metrics.items()
         }
 
diff --git a/src/modalities/loss_functions.py b/src/modalities/loss_functions.py
index e3be6100d..b4cd1367f 100644
--- a/src/modalities/loss_functions.py
+++ b/src/modalities/loss_functions.py
@@ -31,17 +31,29 @@ def __init__(self, target_key: str, prediction_key: str, tag: str = "CLMCrossEnt
         self.prediction_key = prediction_key
         # Mean over the tokens in the local-batch (batch per rank)
         self.loss_fun = CrossEntropyLoss(reduction="mean")
+        self._last_ce_loss = torch.tensor(0.0)
+        self._last_metrics = None
 
     @overload
     def __call__(self, forward_batch: InferenceResultBatch) -> torch.Tensor:
         ...
 
     @overload
-    def __call__(self, outputs: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
+    def __call__(self, outputs: torch.Tensor | dict, targets: torch.Tensor) -> torch.Tensor:
         ...
 
     def __call__(self, *args, **kwargs) -> torch.Tensor:
-        labels, lm_logits = self._parse_arguments(args, kwargs)
+        labels, outputs = self._parse_arguments(args, kwargs)
+
+        if isinstance(outputs, dict):
+            if "logits" in outputs:
+                lm_logits = outputs["logits"]
+            else:
+                lm_logits = outputs[self.prediction_key]
+            metrics = outputs.get("metrics", None)
+        else:
+            lm_logits = outputs
+            metrics = None
 
         # move labels to correct device to enable model parallelism
         labels = labels.to(lm_logits.device)
@@ -49,42 +61,56 @@ def __call__(self, *args, **kwargs) -> torch.Tensor:
         shift_labels = labels.contiguous().long()
         # Flatten the tokens. We compute here, the loss per token.
         loss = self.loss_fun(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+        
+        self._last_ce_loss = loss.detach()
+        self._last_metrics = _detach_metrics(metrics)
         return loss
 
+    def get_metrics(self) -> dict:
+        return {
+            "ce_loss": self._last_ce_loss,
+            "metrics": self._last_metrics,
+        }
+
     def _parse_arguments(
         self,
-        args: list[torch.Tensor] | list[InferenceResultBatch],
-        kwargs: dict[str, torch.Tensor] | dict[str, InferenceResultBatch],
-    ) -> tuple[torch.Tensor, torch.Tensor]:
+        args: list[torch.Tensor | dict] | list[InferenceResultBatch],
+        kwargs: dict[str, torch.Tensor | dict] | dict[str, InferenceResultBatch],
+    ) -> tuple[torch.Tensor, torch.Tensor | dict]:
         if len(args) == 1 and isinstance(args[0], InferenceResultBatch):
             forward_batch = args[0]
             labels = forward_batch.get_targets(self.target_key)
-            lm_logits = forward_batch.get_predictions(self.prediction_key)
+            outputs = forward_batch.predictions
         elif "forward_batch" in kwargs and isinstance(kwargs["forward_batch"], InferenceResultBatch):
             forward_batch = kwargs["forward_batch"]
             labels = forward_batch.get_targets(self.target_key)
-            lm_logits = forward_batch.get_predictions(self.prediction_key)
-        elif len(args) == 2 and all(isinstance(arg, torch.Tensor) for arg in args):
-            lm_logits, labels = args
-        elif (
-            "outputs" in kwargs
-            and "targets" in kwargs
-            and isinstance(kwargs["outputs"], torch.Tensor)
-            and isinstance(kwargs["targets"], torch.Tensor)
-        ):
-            lm_logits = kwargs["outputs"]
+            outputs = forward_batch.predictions
+        elif len(args) == 2:
+            outputs, labels = args
+        elif "outputs" in kwargs and "targets" in kwargs:
+            outputs = kwargs["outputs"]
             labels = kwargs["targets"]
-        elif (
-            len(args) == 1
-            and "targets" in kwargs
-            and isinstance(args[0], torch.Tensor)
-            and isinstance(kwargs["targets"], torch.Tensor)
-        ):
-            lm_logits = args[0]
+        elif len(args) == 1 and "targets" in kwargs:
+            outputs = args[0]
             labels = kwargs["targets"]
         else:
             raise TypeError("Invalid arguments for CLMCrossEntropyLoss.__call__")
-        return labels, lm_logits
+        return labels, outputs
+
+
+def _detach_metrics(metrics: dict | None) -> dict | None:
+    """Recursively detach all tensors in a nested dict."""
+    if metrics is None:
+        return None
+    result = {}
+    for key, value in metrics.items():
+        if isinstance(value, torch.Tensor):
+            result[key] = value.detach()
+        elif isinstance(value, dict):
+            result[key] = _detach_metrics(value)
+        else:
+            result[key] = value
+    return result
 
 
 def nce_loss(
diff --git a/src/modalities/main.py b/src/modalities/main.py
index 49ac97b91..9398b4069 100644
--- a/src/modalities/main.py
+++ b/src/modalities/main.py
@@ -220,6 +220,14 @@ def run(self, components: TrainingComponentsInstantiationModel):
 
         print_rank_0(report)
 
+        conversion_callback = None
+        if components.model_converter is not None:
+            conversion_callback = components.model_converter.convert
+
+        downstream_evaluation_callback = None
+        if components.downstream_evaluator is not None:
+            downstream_evaluation_callback = components.downstream_evaluator.evaluate
+
         gym.run(
             train_data_loader=components.train_dataloader,
             evaluation_data_loaders=components.eval_dataloaders,
@@ -229,8 +237,17 @@ def run(self, components: TrainingComponentsInstantiationModel):
             evaluation_interval_in_steps=components.settings.intervals.evaluation_interval_in_steps,
             training_log_interval_in_steps=components.settings.intervals.training_log_interval_in_steps,
             scheduled_pipeline=components.scheduled_pipeline,
+            conversion_callback=conversion_callback,
+            downstream_evaluation_callback=downstream_evaluation_callback,
         )
 
+        if components.downstream_evaluator is not None:
+            print_rank_0("\n" + "="*80)
+            print_rank_0("Training loop complete! Waiting for background evaluations to finish...")
+            print_rank_0("="*80 + "\n")
+            components.downstream_evaluator.wait_for_evaluations()
+            print_rank_0("All background evaluations completed successfully!")
+
     def get_logging_publishers(
         self,
         progress_subscriber: MessageSubscriberIF[ProgressUpdate],
diff --git a/src/modalities/models/gpt2/gpt2_model.py b/src/modalities/models/gpt2/gpt2_model.py
index f43e6e87b..70e8d3545 100644
--- a/src/modalities/models/gpt2/gpt2_model.py
+++ b/src/modalities/models/gpt2/gpt2_model.py
@@ -1122,7 +1122,7 @@ def __init__(
             )  # https://paperswithcode.com/method/weight-tying
 
     @overload
-    def forward(self, inputs: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
+    def forward(self, inputs: dict[str, torch.Tensor]) -> dict[str, torch.Tensor | dict]:
         """
         Forward pass of the GPT2LLM module.
 
@@ -1131,8 +1131,8 @@ def forward(self, inputs: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
                 - sample_key (str): Key for the input tensor containing token ids.
 
         Returns:
-            dict[str, torch.Tensor]: A dictionary containing output tensors.
-                - prediction_key (str): Key for the output tensor containing logits.
+            dict[str, torch.Tensor | dict]: A dictionary containing output tensors and metrics.
+                - prediction_key (str): Key for the output containing logits and metrics dict.
         """
         ...
 
@@ -1149,7 +1149,7 @@ def forward(self, inputs: torch.Tensor) -> torch.Tensor:
         """
         ...
 
-    def forward(self, inputs: dict[str, torch.Tensor] | torch.Tensor) -> dict[str, torch.Tensor] | torch.Tensor:
+    def forward(self, inputs: dict[str, torch.Tensor] | torch.Tensor) -> dict[str, torch.Tensor | dict] | torch.Tensor:
         """
         Forward pass of the GPT2LLM module.
 
@@ -1157,14 +1157,19 @@ def forward(self, inputs: dict[str, torch.Tensor] | torch.Tensor) -> dict[str, t
             inputs (dict[str, torch.Tensor] | torch.Tensor): Input data.
 
         Returns:
-            dict[str, torch.Tensor] | torch.Tensor: Model output.
+            dict[str, torch.Tensor | dict] | torch.Tensor: Model output.
         """
         if isinstance(inputs, dict):
-            return {self.prediction_key: self.forward_impl(inputs[self.sample_key])}
+            logits, metrics = self.forward_impl(inputs[self.sample_key])
+            return {
+                self.prediction_key: logits,
+                "metrics": metrics,
+            }
         else:
-            return self.forward_impl(inputs)
+            logits, _ = self.forward_impl(inputs)
+            return logits
 
-    def forward_impl(self, inputs: torch.Tensor) -> torch.Tensor:
+    def forward_impl(self, inputs: torch.Tensor) -> tuple[torch.Tensor, dict]:
         """
         Forward pass implementation of the GPT2LLM module.
 
@@ -1172,7 +1177,7 @@ def forward_impl(self, inputs: torch.Tensor) -> torch.Tensor:
             inputs (torch.Tensor): A tensor containing input token ids.
 
         Returns:
-            torch.Tensor: A tensor containing output logits.
+            tuple[torch.Tensor, dict]: A tuple containing output logits and custom metrics.
         """
         device = inputs.device
         seq_len = inputs.size(1)
@@ -1192,11 +1197,31 @@ def forward_impl(self, inputs: torch.Tensor) -> torch.Tensor:
         # TODO: use drop out also without absolute position embedding?
         h = self.transformer.drop(h) if hasattr(self.transformer, "drop") else h
 
+        layer_norms = []
         for layer_idx in self.transformer.h:
             h = self.transformer.h[layer_idx](h)
+            layer_norms.append(h.detach().norm(dim=-1).mean())
+
         h = self.transformer.lm_head_norm(h) if hasattr(self.transformer, "lm_head_norm") else h
-        h = self.transformer.lm_head(h) if hasattr(self.transformer, "lm_head") else h
-        return h
+        logits = self.transformer.lm_head(h) if hasattr(self.transformer, "lm_head") else h
+
+        with torch.no_grad():
+            log_p = torch.log_softmax(logits, dim=-1)
+            p = torch.exp(log_p)
+            entropy = -torch.sum(p * log_p, dim=-1).mean()
+            layer_norms_tensor = torch.stack(layer_norms)
+
+        metrics = {
+            "scalars": {
+                "logits_entropy": entropy
+            },
+            "per_layer_scalars": {
+                "layer_activation_norm": layer_norms_tensor
+            }
+        }
+
+        return logits, metrics
+
 
 
 def manual_scaled_dot_product_attention(
diff --git a/src/modalities/registry/components.py b/src/modalities/registry/components.py
index 26df9b432..3cab47143 100644
--- a/src/modalities/registry/components.py
+++ b/src/modalities/registry/components.py
@@ -33,7 +33,9 @@
     DCPCheckpointSavingConfig,
     DebuggingEnrichedModelConfig,
     DistributedSamplerConfig,
+    DownstreamEvaluatorConfig,
     DummyLRSchedulerConfig,
+    ModelConverterConfig,
     DummyProgressSubscriberConfig,
     DummyResultSubscriberConfig,
     EvaluationResultToDiscSubscriberConfig,
@@ -82,6 +84,8 @@
     ProgressSubscriberFactory,
     ResultsSubscriberFactory,
 )
+from modalities.evaluator import DownstreamEvaluator
+from modalities.conversion.model_converter import ModelConverter
 from modalities.loss_functions import CLMCrossEntropyLoss
 from modalities.models.coca.coca_model import CoCa, CoCaConfig
 from modalities.models.coca.collator import CoCaCollateFnConfig, CoCaCollatorFn
@@ -528,4 +532,6 @@ class ComponentEntity:
         maybe_model_list(HookRegistration.register_print_forward_hooks),
         PrintForwardHookConfig,
     ),
+    ComponentEntity("downstream_evaluator", "default", DownstreamEvaluator, DownstreamEvaluatorConfig),
+    ComponentEntity("model_converter", "default", ModelConverter, ModelConverterConfig),
 ]
diff --git a/src/modalities/trainer.py b/src/modalities/trainer.py
index 4ad54b226..1256707c4 100644
--- a/src/modalities/trainer.py
+++ b/src/modalities/trainer.py
@@ -51,6 +51,10 @@ class ThroughputAggregationKeys(Enum):
     FORWARD_BACKWARD_TIME = "FORWARD_BACKWARD_TIME"
 
 
+
+from modalities.training.logging import MetricsAccumulator, format_metrics
+
+
 class Trainer:
     def __init__(
         self,
@@ -150,7 +154,7 @@ def _train_batch(
                 operate the model. Defaults to None.
 
         Returns:
-            tuple[bool, int, torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]:
+            tuple[bool, int, torch.Tensor, Optional[torch.Tensor]]:
                 A tuple containing the following:
                     - step_performed (bool): Indicates whether a training step was performed.
                     - num_train_steps_done (int): The number of training steps done.
@@ -207,6 +211,8 @@ def train(
         evaluation_callback: Callable[[int], None],
         checkpointing_callback: Callable[[TrainingProgress], None],
         scheduled_pipeline: Pipeline | None = None,
+        conversion_callback: Optional[Callable[[int], None]] = None,
+        downstream_evaluation_callback: Optional[Callable[[int], None]] = None,
     ):
         """
         Trains the model.
@@ -220,6 +226,8 @@ def train(
             checkpointing_callback (Callable[[TrainingProgress], None]): A callback function for checkpointing.
             scheduled_pipeline (Pipeline | None, optional): In case of pipeline parallelism, this is used to
                 operate the model. Defaults to None.
+            conversion_callback (Optional[Callable[[int], None]]): A callback function for checkpoint-to-HF conversion.
+            downstream_evaluation_callback (Optional[Callable[[int], None]]): A callback function for downstream evaluation.
 
         Returns:
             None
@@ -234,6 +242,7 @@ def train(
 
         local_num_seen_samples = 0
         cumulated_losses = torch.zeros(3).cuda()
+        metrics_accum = MetricsAccumulator()
 
         # throughput
         device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
@@ -257,6 +266,10 @@ def train(
             num_target_tokens=self.num_target_tokens,
         )
         checkpointing_callback(training_progress=training_progress)
+        if conversion_callback is not None:
+            conversion_callback(num_train_steps_done=self.num_seen_train_steps)
+        if downstream_evaluation_callback is not None:
+            downstream_evaluation_callback(num_train_steps_done=self.num_seen_train_steps)
 
         num_steps_todo = self.num_target_steps - self.num_seen_train_steps
         num_batches_todo = num_steps_todo * self.gradient_acc_steps
@@ -291,6 +304,9 @@ def train(
                     # it has less samples than the batch size
                     cumulated_losses[-1] += 1  # number of local batches
 
+                    if hasattr(loss_fun, "get_metrics"):
+                        metrics_accum.accumulate(loss_fun.get_metrics())
+
                 # gradient norm is already synced across all ranks
                 if gradient_norm_score is not None:
                     gradient_norm_scores.append(gradient_norm_score.item())
@@ -336,17 +352,52 @@ def train(
                         reduced_losses[0],
                         reduced_losses[1],
                     )
+
+                    adaptive_losses = {}
+                    adaptive_metrics = {}
+                    if metrics_accum.count > 0:
+                        (
+                            sync_tensor, scalar_names, per_layer_names, per_layer_sizes,
+                            hist_names, hist_shapes,
+                        ) = metrics_accum.build_sync_tensor(device)
+
+                        reduce_scale = dist.get_world_size() / self.pp_degree
+                        synced_tensor = Reducer.reduce(
+                            tensor=sync_tensor,
+                            operation=dist.ReduceOp.SUM,
+                            post_processing_fun=lambda t: t / reduce_scale,
+                        )
+
+                        (
+                            synced_loss, synced_scalars, synced_per_layer, synced_hists,
+                        ) = MetricsAccumulator.unpack_synced_tensor(
+                            synced_tensor, scalar_names, per_layer_names, per_layer_sizes,
+                            hist_names, hist_shapes,
+                        )
+
+                        adaptive_losses, adaptive_metrics = format_metrics(
+                            loss=synced_loss,
+                            scalars=synced_scalars,
+                            per_layer_scalars=synced_per_layer,
+                            per_layer_vectors=metrics_accum.last_per_layer_vectors,
+                            per_layer_histograms=synced_hists,
+                        )
+
                     losses = {
                         "train loss avg": ResultItem(train_loss_avg, decimal_places=2),
                         "train loss last": ResultItem(train_loss_last_batch, decimal_places=2),
+                        **adaptive_losses,
                     }
 
                     metrics = {
                         "consumed tokens": ResultItem(torch.tensor(training_progress.num_seen_tokens_total), 0),
                         "grad norm avg": ResultItem(torch.mean(torch.Tensor(gradient_norm_scores)), 2),
                         "grad norm last": ResultItem(torch.tensor(gradient_norm_scores[-1]), 2),
+                        **adaptive_metrics,
                     }
+
                     gradient_norm_scores = []
+
                     mfu_score = torch.tensor(-1.0)
                     if self.mfu_calculator is not None:
                         mfu_score = self.mfu_calculator.compute(num_samples_per_second=global_num_samples_per_second)
@@ -384,10 +435,15 @@ def train(
                     )
 
                     cumulated_losses.zero_()
+                    metrics_accum.reset()
                 if step_performed:
                     self.gc.run(step_count=training_progress.num_seen_steps_total)
                     evaluation_callback(num_train_steps_done=training_progress.num_seen_steps_total)
                     checkpointing_callback(training_progress=training_progress)
+                    if conversion_callback is not None:
+                        conversion_callback(num_train_steps_done=training_progress.num_seen_steps_total)
+                    if downstream_evaluation_callback is not None:
+                        downstream_evaluation_callback(num_train_steps_done=training_progress.num_seen_steps_total)
 
                 profiler_cm.step()
 
diff --git a/src/modalities/training/logging.py b/src/modalities/training/logging.py
new file mode 100644
index 000000000..18b632af9
--- /dev/null
+++ b/src/modalities/training/logging.py
@@ -0,0 +1,198 @@
+from typing import Optional
+import torch
+from modalities.batch import ResultItem
+
+
+# =============================================================================
+# Generic Metrics Accumulator
+# =============================================================================
+
+class MetricsAccumulator:
+    """Accumulates metrics across batches and produces a single flat tensor
+    for cross-rank reduction.
+    """
+
+    def __init__(self):
+        self.reset()
+
+    def reset(self):
+        self.loss_sum: float = 0.0
+        self.scalar_sums: dict[str, float] = {}
+        self.per_layer_scalar_sums: dict[str, torch.Tensor] = {}
+        self.last_per_layer_vectors: dict[str, torch.Tensor] = {}
+        self.per_layer_hist_sums: dict[str, torch.Tensor] = {}
+        self.count: int = 0
+
+    def accumulate(self, loss_metrics: dict):
+        if "ce_loss" in loss_metrics:
+            self.loss_sum += loss_metrics["ce_loss"].item()
+        elif "loss" in loss_metrics:
+            self.loss_sum += loss_metrics["loss"].item()
+        self.count += 1
+
+        bag = loss_metrics.get("metrics")
+        if bag is None:
+            return
+
+        for name, tensor in bag.get("scalars", {}).items():
+            self.scalar_sums[name] = self.scalar_sums.get(name, 0.0) + tensor.item()
+
+        for name, tensor in bag.get("per_layer_scalars", {}).items():
+            if name not in self.per_layer_scalar_sums:
+                self.per_layer_scalar_sums[name] = torch.zeros_like(tensor, dtype=torch.float32)
+            self.per_layer_scalar_sums[name] += tensor.float()
+
+        for name, tensor in bag.get("per_layer_vectors", {}).items():
+            self.last_per_layer_vectors[name] = tensor
+
+        for name, tensor in bag.get("per_layer_histograms", {}).items():
+            if name not in self.per_layer_hist_sums:
+                self.per_layer_hist_sums[name] = torch.zeros_like(tensor, dtype=torch.float32)
+            self.per_layer_hist_sums[name] += tensor.float()
+
+    def build_sync_tensor(
+        self, device: torch.device
+    ) -> tuple[
+        torch.Tensor,
+        list[str],
+        list[str],
+        dict[str, int],
+        list[str],
+        dict[str, tuple],
+    ]:
+        if self.count == 0:
+            return torch.zeros(1, device=device), [], [], {}, [], {}
+
+        n = self.count
+        values = [self.loss_sum / n]
+
+        scalar_names = sorted(self.scalar_sums.keys())
+        for name in scalar_names:
+            values.append(self.scalar_sums[name] / n)
+
+        per_layer_names = sorted(self.per_layer_scalar_sums.keys())
+        per_layer_sizes = {}
+        layer_tensors = []
+        for name in per_layer_names:
+            t = self.per_layer_scalar_sums[name] / n
+            layer_tensors.append(t.to(device))
+            per_layer_sizes[name] = t.numel()
+
+        hist_names = sorted(self.per_layer_hist_sums.keys())
+        hist_shapes: dict[str, tuple] = {}
+        hist_tensors = []
+        for name in hist_names:
+            t = self.per_layer_hist_sums[name] / n
+            hist_shapes[name] = tuple(t.shape)
+            hist_tensors.append(t.to(device).flatten())
+
+        combined = torch.tensor(values, device=device, dtype=torch.float32)
+        if layer_tensors:
+            combined = torch.cat([combined, torch.cat(layer_tensors)])
+        if hist_tensors:
+            combined = torch.cat([combined, torch.cat(hist_tensors)])
+
+        return combined, scalar_names, per_layer_names, per_layer_sizes, hist_names, hist_shapes
+
+    @staticmethod
+    def unpack_synced_tensor(
+        synced: torch.Tensor,
+        scalar_names: list[str],
+        per_layer_names: list[str],
+        per_layer_sizes: dict[str, int],
+        hist_names: list[str] = None,
+        hist_shapes: dict[str, tuple] = None,
+    ) -> tuple[
+        torch.Tensor,
+        dict[str, torch.Tensor],
+        dict[str, torch.Tensor],
+        dict[str, torch.Tensor],
+    ]:
+        hist_names = hist_names or []
+        hist_shapes = hist_shapes or {}
+
+        idx = 0
+        loss = synced[idx]; idx += 1
+
+        scalars = {}
+        for name in scalar_names:
+            scalars[name] = synced[idx]; idx += 1
+
+        per_layer_scalars = {}
+        for name in per_layer_names:
+            size = per_layer_sizes[name]
+            per_layer_scalars[name] = synced[idx : idx + size]; idx += size
+
+        per_layer_histograms = {}
+        for name in hist_names:
+            shape = hist_shapes[name]
+            size = 1
+            for dim in shape:
+                size *= dim
+            per_layer_histograms[name] = synced[idx : idx + size].reshape(shape)
+            idx += size
+
+        return loss, scalars, per_layer_scalars, per_layer_histograms
+
+
+# =============================================================================
+# Metrics Formatter
+# =============================================================================
+
+def format_metrics(
+    loss: torch.Tensor,
+    scalars: dict[str, torch.Tensor],
+    per_layer_scalars: dict[str, torch.Tensor],
+    per_layer_vectors: dict[str, torch.Tensor],
+    summary_only: bool = False,
+    per_layer_histograms: Optional[dict[str, torch.Tensor]] = None,
+) -> tuple[dict[str, ResultItem], dict[str, ResultItem]]:
+    per_layer_histograms = per_layer_histograms or {}
+
+    losses = {
+        "loss/ce_avg": ResultItem(loss, decimal_places=2),
+    }
+
+    metrics: dict[str, ResultItem] = {}
+
+    for name, val in scalars.items():
+        metrics[f"adaptive/{name}"] = ResultItem(val, 4)
+
+    for name, vals in per_layer_scalars.items():
+        metrics[f"summary/{name}"] = ResultItem(vals.mean(), 4)
+        if not summary_only:
+            for i, v in enumerate(vals):
+                metrics[f"layer_{i}/{name}"] = ResultItem(v, 4)
+
+    for name, tensor in per_layer_vectors.items():
+        if tensor.numel() == 0:
+            continue
+        t = tensor.float().cpu()
+        n_layers, n_loops = t.shape
+
+        metrics[f"summary/{name}"] = ResultItem(t.mean(), 4)
+
+        if not summary_only:
+            for i in range(n_layers):
+                metrics[f"layer_{i}/avg_{name}"] = ResultItem(t[i].mean(), 4)
+                for j in range(n_loops):
+                    metrics[f"layer_{i}/{name}_{j}"] = ResultItem(t[i, j], 4)
+
+            for j in range(n_loops):
+                metrics[f"loop_{j}/{name}"] = ResultItem(t[:, j].mean(), 4)
+
+    for name, tensor in per_layer_histograms.items():
+        if tensor.numel() == 0:
+            continue
+        t = tensor.float().cpu()
+        n_layers, n_bins = t.shape
+
+        for b in range(n_bins):
+            metrics[f"hist/{name}/bin_{b}"] = ResultItem(t[:, b].mean(), 4)
+
+        if not summary_only:
+            for i in range(n_layers):
+                for b in range(n_bins):
+                    metrics[f"hist/{name}/layer_{i}/bin_{b}"] = ResultItem(t[i, b], 4)
+
+    return losses, metrics
diff --git a/tests/logging_broker/subscriber_impl/test_logging_rich.py b/tests/logging_broker/subscriber_impl/test_logging_rich.py
new file mode 100644
index 000000000..89710fc2b
--- /dev/null
+++ b/tests/logging_broker/subscriber_impl/test_logging_rich.py
@@ -0,0 +1,120 @@
+import torch
+import pytest
+from modalities.training.logging import MetricsAccumulator, format_metrics
+from modalities.batch import ResultItem
+
+
+def test_metrics_accumulator_accumulation_and_sync():
+    device = torch.device("cpu")
+    accum = MetricsAccumulator()
+
+    # Step 1: Accumulate first batch
+    metrics_1 = {
+        "scalars": {
+            "p_weight": torch.tensor(0.5)
+        },
+        "per_layer_scalars": {
+            "cost": torch.tensor([1.0, 2.0])
+        },
+        "per_layer_vectors": {
+            "vec": torch.tensor([[0.1, 0.2], [0.3, 0.4]])
+        },
+        "per_layer_histograms": {
+            "hist": torch.tensor([[0.5, 0.5], [0.8, 0.2]])
+        }
+    }
+    accum.accumulate({"ce_loss": torch.tensor(2.0), "metrics": metrics_1})
+
+    # Step 2: Accumulate second batch
+    metrics_2 = {
+        "scalars": {
+            "p_weight": torch.tensor(1.5)
+        },
+        "per_layer_scalars": {
+            "cost": torch.tensor([3.0, 4.0])
+        },
+        "per_layer_vectors": {
+            # Vectors are last-batch-only in trainer design
+            "vec": torch.tensor([[1.1, 1.2], [1.3, 1.4]])
+        },
+        "per_layer_histograms": {
+            "hist": torch.tensor([[0.3, 0.7], [0.6, 0.4]])
+        }
+    }
+    accum.accumulate({"ce_loss": torch.tensor(4.0), "metrics": metrics_2})
+
+    assert accum.count == 2
+
+    # Step 3: Build sync tensor
+    sync_tensor, scalar_names, pl_names, pl_sizes, hist_names, hist_shapes = accum.build_sync_tensor(device)
+
+    # Expected averages:
+    # ce_loss = (2 + 4) / 2 = 3.0
+    # scalars: p_weight = (0.5 + 1.5) / 2 = 1.0
+    # per_layer_scalars: cost = [(1+3)/2, (2+4)/2] = [2.0, 3.0]
+    # per_layer_histograms: hist = [[(0.5+0.3)/2, (0.5+0.7)/2], [(0.8+0.6)/2, (0.2+0.4)/2]] = [[0.4, 0.6], [0.7, 0.3]]
+
+    # Step 4: Unpack
+    loss, scalars, pl_scalars, pl_hist = MetricsAccumulator.unpack_synced_tensor(
+        sync_tensor, scalar_names, pl_names, pl_sizes, hist_names, hist_shapes
+    )
+
+    assert torch.allclose(loss, torch.tensor(3.0))
+    assert torch.allclose(scalars["p_weight"], torch.tensor(1.0))
+    assert torch.allclose(pl_scalars["cost"], torch.tensor([2.0, 3.0]))
+    assert torch.allclose(pl_hist["hist"], torch.tensor([[0.4, 0.6], [0.7, 0.3]]))
+    assert torch.allclose(accum.last_per_layer_vectors["vec"], torch.tensor([[1.1, 1.2], [1.3, 1.4]]))
+
+
+def test_format_metrics():
+    loss = torch.tensor(3.0)
+    scalars = {"p_weight": torch.tensor(1.0)}
+    pl_scalars = {"cost": torch.tensor([2.0, 3.0])}
+    pl_vectors = {"vec": torch.tensor([[1.1, 1.2], [1.3, 1.4]])}
+    pl_hists = {"hist": torch.tensor([[0.4, 0.6], [0.7, 0.3]])}
+
+    # Test summary_only = False
+    losses, metrics = format_metrics(
+        loss=loss,
+        scalars=scalars,
+        per_layer_scalars=pl_scalars,
+        per_layer_vectors=pl_vectors,
+        summary_only=False,
+        per_layer_histograms=pl_hists
+    )
+
+    assert losses["loss/ce_avg"].value.item() == pytest.approx(3.0)
+    assert metrics["adaptive/p_weight"].value.item() == pytest.approx(1.0)
+    assert metrics["summary/cost"].value.item() == pytest.approx(2.5)
+    assert metrics["layer_0/cost"].value.item() == pytest.approx(2.0)
+    assert metrics["layer_1/cost"].value.item() == pytest.approx(3.0)
+
+    # Vectors
+    assert metrics["summary/vec"].value.item() == pytest.approx(1.25)
+    assert metrics["layer_0/vec_0"].value.item() == pytest.approx(1.1)
+    assert metrics["layer_0/vec_1"].value.item() == pytest.approx(1.2)
+    assert metrics["layer_1/vec_0"].value.item() == pytest.approx(1.3)
+    assert metrics["layer_1/vec_1"].value.item() == pytest.approx(1.4)
+
+    # Histograms
+    assert metrics["hist/hist/bin_0"].value.item() == pytest.approx(0.55)
+    assert metrics["hist/hist/bin_1"].value.item() == pytest.approx(0.45)
+    assert metrics["hist/hist/layer_0/bin_0"].value.item() == pytest.approx(0.4)
+    assert metrics["hist/hist/layer_1/bin_1"].value.item() == pytest.approx(0.3)
+
+    # Test summary_only = True
+    losses, metrics = format_metrics(
+        loss=loss,
+        scalars=scalars,
+        per_layer_scalars=pl_scalars,
+        per_layer_vectors=pl_vectors,
+        summary_only=True,
+        per_layer_histograms=pl_hists
+    )
+
+    assert "layer_0/cost" not in metrics
+    assert "layer_0/vec_0" not in metrics
+    assert "hist/hist/layer_0/bin_0" not in metrics
+    assert metrics["summary/cost"].value.item() == pytest.approx(2.5)
+    assert metrics["summary/vec"].value.item() == pytest.approx(1.25)
+    assert metrics["hist/hist/bin_0"].value.item() == pytest.approx(0.55)
diff --git a/tests/test_downstream_evaluator.py b/tests/test_downstream_evaluator.py
new file mode 100644
index 000000000..ffd535d03
--- /dev/null
+++ b/tests/test_downstream_evaluator.py
@@ -0,0 +1,281 @@
+import json
+import os
+import tempfile
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+import torch.nn as nn
+from pydantic import BaseModel
+
+from modalities.config.component_factory import ComponentFactory
+from modalities.conversion.model_converter import ModelConverter
+from modalities.evaluator import DownstreamEvaluator
+from modalities.registry.components import ComponentEntity
+from modalities.registry.registry import Registry
+from modalities.tokenization.tokenizer_wrapper import TokenizerWrapper
+
+
+# ---------- helpers ----------
+
+class MockTokenizer(TokenizerWrapper):
+    def tokenize(self, text: str) -> list[int]:
+        return []
+
+    def decode(self, input_ids: list[int]) -> str:
+        return ""
+
+    @property
+    def vocab_size(self) -> int:
+        return 0
+
+    def get_token_id(self, token: str) -> int:
+        return 0
+
+    def is_special_token_id(self, token_id: int) -> bool:
+        return False
+
+
+# ---------- ModelConverter tests ----------
+
+def test_model_converter_skips_non_matching_step():
+    converter = ModelConverter(
+        command_template="echo {checkpoint_path} {output_dir}",
+        checkpoint_dir=Path("/tmp/fake"),
+        global_rank=0,
+        eval_interval=5,
+    )
+    with patch("subprocess.run") as mock_run:
+        converter.convert(num_train_steps_done=3)
+        mock_run.assert_not_called()
+
+
+def test_model_converter_skips_step_zero():
+    converter = ModelConverter(
+        command_template="echo {checkpoint_path} {output_dir}",
+        checkpoint_dir=Path("/tmp/fake"),
+        global_rank=0,
+        eval_interval=5,
+    )
+    with patch("subprocess.run") as mock_run:
+        converter.convert(num_train_steps_done=0)
+        mock_run.assert_not_called()
+
+
+def test_model_converter_skips_non_rank_zero():
+    converter = ModelConverter(
+        command_template="echo {checkpoint_path} {output_dir}",
+        checkpoint_dir=Path("/tmp/fake"),
+        global_rank=1,
+        eval_interval=5,
+    )
+    with patch("subprocess.run") as mock_run:
+        converter.convert(num_train_steps_done=5)
+        mock_run.assert_not_called()
+
+
+def test_model_converter_runs_command_on_matching_step():
+    with tempfile.TemporaryDirectory() as tmpdir:
+        checkpoint_dir = Path(tmpdir)
+        ckpt_path = checkpoint_dir / "step_5"
+        ckpt_path.mkdir()
+
+        info = {"checkpoint_folder_path": str(ckpt_path)}
+        with open(checkpoint_dir / "last_checkpoint_info.json", "w") as f:
+            json.dump(info, f)
+
+        converter = ModelConverter(
+            command_template="echo {checkpoint_path} {output_dir}",
+            checkpoint_dir=checkpoint_dir,
+            global_rank=0,
+            eval_interval=5,
+        )
+
+        with patch("subprocess.run") as mock_run:
+            converter.convert(num_train_steps_done=5)
+            mock_run.assert_called_once()
+            cmd = mock_run.call_args[0][0]
+            assert str(ckpt_path) in cmd
+            assert "hf_checkpoint" in cmd
+
+
+def test_model_converter_skips_if_hf_checkpoint_exists():
+    with tempfile.TemporaryDirectory() as tmpdir:
+        checkpoint_dir = Path(tmpdir)
+        ckpt_path = checkpoint_dir / "step_5"
+        ckpt_path.mkdir()
+        (ckpt_path / "hf_checkpoint").mkdir()
+
+        info = {"checkpoint_folder_path": str(ckpt_path)}
+        with open(checkpoint_dir / "last_checkpoint_info.json", "w") as f:
+            json.dump(info, f)
+
+        converter = ModelConverter(
+            command_template="echo {checkpoint_path} {output_dir}",
+            checkpoint_dir=checkpoint_dir,
+            global_rank=0,
+            eval_interval=5,
+        )
+
+        with patch("subprocess.run") as mock_run:
+            converter.convert(num_train_steps_done=5)
+            mock_run.assert_not_called()
+
+
+# ---------- DownstreamEvaluator tests ----------
+
+def test_downstream_evaluator_skips_non_matching_step():
+    evaluator = DownstreamEvaluator(
+        tokenizer=MockTokenizer(),
+        tasks=["arc_challenge::olmes"],
+        eval_interval=5,
+        checkpoint_dir=Path("/tmp/fake"),
+        global_rank=0,
+        olmes_command_template="echo {hf_model_dir} {tasks} {step}",
+    )
+    with patch("subprocess.Popen") as mock_popen:
+        evaluator.evaluate(num_train_steps_done=3)
+        mock_popen.assert_not_called()
+
+
+def test_downstream_evaluator_skips_non_rank_zero():
+    evaluator = DownstreamEvaluator(
+        tokenizer=MockTokenizer(),
+        tasks=["arc_challenge::olmes"],
+        eval_interval=5,
+        checkpoint_dir=Path("/tmp/fake"),
+        global_rank=1,
+        olmes_command_template="echo {hf_model_dir} {tasks} {step}",
+    )
+    with patch("subprocess.Popen") as mock_popen:
+        evaluator.evaluate(num_train_steps_done=5)
+        mock_popen.assert_not_called()
+
+
+def test_downstream_evaluator_runs_when_hf_checkpoint_exists():
+    with tempfile.TemporaryDirectory() as tmpdir:
+        checkpoint_dir = Path(tmpdir)
+        ckpt_path = checkpoint_dir / "step_10"
+        ckpt_path.mkdir()
+        hf_dir = ckpt_path / "hf_checkpoint"
+        hf_dir.mkdir()
+
+        info = {"checkpoint_folder_path": str(ckpt_path)}
+        with open(checkpoint_dir / "last_checkpoint_info.json", "w") as f:
+            json.dump(info, f)
+
+        evaluator = DownstreamEvaluator(
+            tokenizer=MockTokenizer(),
+            tasks=["arc_challenge::olmes", "hellaswag::olmes"],
+            eval_interval=10,
+            checkpoint_dir=checkpoint_dir,
+            global_rank=0,
+            olmes_command_template="olmes --model {hf_model_dir} --tasks {tasks} --step {step}",
+        )
+
+        with patch("subprocess.Popen") as mock_popen:
+            evaluator.evaluate(num_train_steps_done=10)
+            mock_popen.assert_called_once()
+            cmd = mock_popen.call_args[0][0]
+            assert str(hf_dir) in cmd
+            assert "arc_challenge::olmes,hellaswag::olmes" in cmd
+            assert "10" in cmd
+
+
+def test_downstream_evaluator_skips_when_no_hf_checkpoint():
+    with tempfile.TemporaryDirectory() as tmpdir:
+        checkpoint_dir = Path(tmpdir)
+        ckpt_path = checkpoint_dir / "step_10"
+        ckpt_path.mkdir()
+        # No hf_checkpoint folder
+
+        info = {"checkpoint_folder_path": str(ckpt_path)}
+        with open(checkpoint_dir / "last_checkpoint_info.json", "w") as f:
+            json.dump(info, f)
+
+        evaluator = DownstreamEvaluator(
+            tokenizer=MockTokenizer(),
+            tasks=["arc_challenge::olmes"],
+            eval_interval=10,
+            checkpoint_dir=checkpoint_dir,
+            global_rank=0,
+            olmes_command_template="echo {hf_model_dir} {tasks} {step}",
+        )
+
+        with patch("subprocess.Popen") as mock_popen:
+            evaluator.evaluate(num_train_steps_done=10)
+            mock_popen.assert_not_called()
+
+
+# ---------- Factory instantiation tests ----------
+
+def test_downstream_evaluator_factory_instantiation():
+    from modalities.config.config import DownstreamEvaluatorConfig
+    from modalities.config.pydantic_if_types import PydanticDownstreamEvaluatorType
+    from modalities.registry.components import COMPONENTS
+
+    registry = Registry(COMPONENTS)
+    component_factory = ComponentFactory(registry=registry)
+
+    tokenizer_mock = MockTokenizer()
+
+    class TrainingModel(BaseModel):
+        downstream_eval: PydanticDownstreamEvaluatorType
+
+    config_dict = {
+        "downstream_eval": {
+            "component_key": "downstream_evaluator",
+            "variant_key": "default",
+            "config": {
+                "tokenizer": tokenizer_mock,
+                "tasks": ["task_a"],
+                "eval_interval": 10,
+                "checkpoint_dir": "/tmp/test_checkpoints",
+                "global_rank": 0,
+                "olmes_command_template": "echo {hf_model_dir}",
+            },
+        }
+    }
+
+    components = component_factory.build_components(
+        config_dict=config_dict,
+        components_model_type=TrainingModel,
+    )
+
+    assert isinstance(components.downstream_eval, DownstreamEvaluator)
+    assert components.downstream_eval.tokenizer == tokenizer_mock
+    assert components.downstream_eval.tasks == ["task_a"]
+    assert components.downstream_eval.eval_interval == 10
+
+
+def test_model_converter_factory_instantiation():
+    from modalities.config.config import ModelConverterConfig
+    from modalities.config.pydantic_if_types import PydanticModelConverterType
+    from modalities.registry.components import COMPONENTS
+
+    registry = Registry(COMPONENTS)
+    component_factory = ComponentFactory(registry=registry)
+
+    class ConverterModel(BaseModel):
+        converter: PydanticModelConverterType
+
+    config_dict = {
+        "converter": {
+            "component_key": "model_converter",
+            "variant_key": "default",
+            "config": {
+                "command_template": "echo {checkpoint_path} {output_dir}",
+                "checkpoint_dir": "/tmp/test_checkpoints",
+                "global_rank": 0,
+                "eval_interval": 100,
+            },
+        }
+    }
+
+    components = component_factory.build_components(
+        config_dict=config_dict,
+        components_model_type=ConverterModel,
+    )
+
+    assert isinstance(components.converter, ModelConverter)
+    assert components.converter.eval_interval == 100